From df46097605b324f8d880230d960aa5604f916e93 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Tue, 3 Nov 2026 22:02:25 +0530
Subject: [PATCH 01/12] feat(local): draw the dashboard's frame while startup
 loads

Replaces the logo loader. Once onboarding has been seen done on this
database, startup draws the dashboard's columns, headings and composer
box with the dashboard's own components, then fades the contents in.
First launches, onboarding and Docker keep the plain background.
---
 docs/architecture/chat.md                     |   2 +-
 docs/architecture/overview.md                 |   1 +
 surfsense_local/electron/src/main/index.ts    |   6 +
 .../startup-frame/database-identity.test.ts   |  42 ++++
 .../main/startup-frame/database-identity.ts   |  20 ++
 .../main/startup-frame/startup-frame-ipc.ts   |  16 ++
 surfsense_local/electron/src/preload/index.ts |   3 +
 .../src/app/app-bootstrap-startup.test.tsx    | 222 ++++++++++++++++++
 .../frontend/src/app/app-bootstrap.test.tsx   |  18 --
 .../frontend/src/app/app-bootstrap.tsx        | 126 ++++++----
 .../frontend/src/app/logo-fill-loader.tsx     |  55 -----
 .../src/features/chat/chat-composer.tsx       |  32 +--
 .../src/features/chat/chat-viewport-layout.ts |  19 ++
 .../src/features/chat/chat-viewport.tsx       |  19 +-
 .../src/features/chat/composer-layout.ts      |  33 +++
 .../src/features/chat/composer-outline.tsx    |  36 +++
 .../src/features/chat/conversation-column.tsx |  39 +++
 .../frontend/src/features/chat/last-thread.ts |  27 +++
 .../src/features/chat/messages-skeleton.tsx   |  17 ++
 .../src/features/chat/thread-panel.tsx        |  73 ++----
 .../src/features/chat/thread-welcome.tsx      |  13 +
 .../src/features/chat/use-chat-runtime.ts     |  31 +--
 .../features/dashboard/dashboard-columns.tsx  |   9 +-
 .../src/features/dashboard/dashboard-page.tsx |  15 +-
 .../features/dashboard/dashboard-shell.tsx    |  22 ++
 .../src/features/dashboard/left-sidebar.tsx   |  31 ++-
 .../src/features/dashboard/right-panel.tsx    |  12 +-
 .../src/features/dashboard/sidebar-frame.tsx  |  29 +++
 .../features/startup/dashboard-hint.test.ts   |  56 +++++
 .../src/features/startup/dashboard-hint.ts    |  45 ++++
 .../src/features/startup/startup-frame.tsx    |  77 ++++++
 .../features/startup/startup-layout.test.ts   |  32 +++
 .../src/features/startup/startup-layout.ts    |  24 ++
 .../startup/startup-reveal-context.ts         |   4 +
 .../startup/startup-reveal-provider.tsx       |  28 +++
 .../features/startup/startup-reveal.test.tsx  |  73 ++++++
 .../startup/startup-screen-imports.test.ts    |  77 ++++++
 .../src/features/startup/startup-screen.tsx   |  25 ++
 .../features/startup/use-startup-reveal.ts    |  21 ++
 .../src/features/updates/stub-bridge.ts       |   1 +
 .../src/features/workspaces/last-workspace.ts |  19 ++
 .../src/features/workspaces/use-workspaces.ts |  21 +-
 .../features/workspaces/workspace-rail.tsx    |   7 +-
 surfsense_local/frontend/src/index.css        |  49 +---
 surfsense_local/frontend/src/lib/api.ts       |   2 +
 45 files changed, 1232 insertions(+), 297 deletions(-)
 create mode 100644 surfsense_local/electron/src/main/startup-frame/database-identity.test.ts
 create mode 100644 surfsense_local/electron/src/main/startup-frame/database-identity.ts
 create mode 100644 surfsense_local/electron/src/main/startup-frame/startup-frame-ipc.ts
 create mode 100644 surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
 delete mode 100644 surfsense_local/frontend/src/app/logo-fill-loader.tsx
 create mode 100644 surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts
 create mode 100644 surfsense_local/frontend/src/features/chat/composer-layout.ts
 create mode 100644 surfsense_local/frontend/src/features/chat/composer-outline.tsx
 create mode 100644 surfsense_local/frontend/src/features/chat/conversation-column.tsx
 create mode 100644 surfsense_local/frontend/src/features/chat/last-thread.ts
 create mode 100644 surfsense_local/frontend/src/features/chat/messages-skeleton.tsx
 create mode 100644 surfsense_local/frontend/src/features/chat/thread-welcome.tsx
 create mode 100644 surfsense_local/frontend/src/features/dashboard/dashboard-shell.tsx
 create mode 100644 surfsense_local/frontend/src/features/dashboard/sidebar-frame.tsx
 create mode 100644 surfsense_local/frontend/src/features/startup/dashboard-hint.test.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/dashboard-hint.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-frame.tsx
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-layout.test.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-layout.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-reveal-context.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-reveal-provider.tsx
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-screen.tsx
 create mode 100644 surfsense_local/frontend/src/features/startup/use-startup-reveal.ts
 create mode 100644 surfsense_local/frontend/src/features/workspaces/last-workspace.ts

diff --git a/docs/architecture/chat.md b/docs/architecture/chat.md
index b55f54ca1e..4a22bdc0f4 100644
--- a/docs/architecture/chat.md
+++ b/docs/architecture/chat.md
@@ -262,7 +262,7 @@ Interface strings are `dictation_<surface>_<purpose>`, and a server code is `dic
 - The composer follows assistant-ui's keyboard behaviour.
 - While the turn waits, the thinking header's `role="status"` region says "Thinking" or the reading quarter. A second region (`reply-announcer.tsx`) stays silent until it says "Reply finished" once, for a reply it watched run to completion. It is keyed by message, because messages render by index and an instance would otherwise watch one thread's reply and announce another's. Tokens are never fed to it, since a polite region read on every delta would interrupt without pause; a reply loaded from history, a turn that ended with no answer, and one that failed or was stopped after text arrived say nothing; its alert, or the stop, already does. A stopped reply is marked `incomplete` with reason `cancelled` by the runtime ([`use-chat-runtime.ts`](../../surfsense_local/frontend/src/features/chat/use-chat-runtime.ts)), since assistant-ui would otherwise call it complete once the run ends. A queued reply's header region says "Waiting for another reply" with its place. The Chats button's count is in its accessible name ("Chats, 2 replies being written"), and an unread row carries "New reply" as screen-reader text.
 - The conversation has a heading, the thread's title, hidden visually because the visible title is the rename button. Focus does not move to it: on every thread switch the composer keeps the focus assistant-ui gives it, so someone can switch and type, and the composer is described by the heading (`aria-describedby`), so a screen reader names the conversation it writes into. A new chat has no heading until it is a thread.
-- The startup loader and the typed-in thread title respect `prefers-reduced-motion`.
+- The startup fade and the typed-in thread title respect `prefers-reduced-motion`.
 - The mic carries `aria-pressed` and a name that follows its state: "Start voice input", "Stop voice input", or "Set up voice input"; its tooltip is visual only. While voice input runs, a `role="status"` line in the strip says "Starting the microphone…", "Listening", "Loading the speech recognition model…", "Transcribing…" or the 30-second warning; a problem is a `role="alert"`, the two limits a `role="status"`. Under `prefers-reduced-motion` the level bars give way to the still dot beside them.
 
 ## Agent threads
diff --git a/docs/architecture/overview.md b/docs/architecture/overview.md
index 14ae231806..dd984af390 100644
--- a/docs/architecture/overview.md
+++ b/docs/architecture/overview.md
@@ -110,6 +110,7 @@ The rules that keep the processes out of each other's way:
 - API calls go through `request()` in [`lib/api.ts`](../../surfsense_local/frontend/src/lib/api.ts), except the Studio viewers, which fetch an artifact's file bytes directly; an `<img>`, `<audio>` or download link takes its absolute address from `apiUrl()`. `request()` prefixes the address, turns an error body into an `ApiError` with its `code`, and when a user action is refused with `403 egress_disabled` it asks for consent and retries once.
 - TanStack Query holds most server state: threads and messages, the citation panel, the model catalog and settings, the Studio formats and artifact list, and an open artifact with what its viewer reads. An `artifacts` event invalidates the list and every open artifact under one key prefix ([`features/studio/query-keys.ts`](../../surfsense_local/frontend/src/features/studio/query-keys.ts)), so a regenerated artifact that is open shows its new generation. The sources list is component state, reloaded on the events above.
 - There is no router. [`app/app-bootstrap.tsx`](../../surfsense_local/frontend/src/app/app-bootstrap.tsx) asks `GET /llm/onboarding`, then renders either onboarding ([`local-models/selection.md`](local-models/selection.md)) or the lazily loaded dashboard: the workspace rail, threads and sources on the left, the conversation in the middle, and Studio, artifacts and the citation panel on the right. The left and right columns resize by dragging their inner edge, and the conversation keeps at least 520 px ([studio](studio.md#the-studio-panel)).
+- While startup waits on the API, [`features/startup/`](../../surfsense_local/frontend/src/features/startup/) draws the dashboard's constant parts, with nothing to press: the columns at their saved widths, the "SurfSense" and "Studio" headings, and the composer's empty box where the last thread put it. The dashboard draws the same parts with the same components, so nothing moves when it takes over; its contents fade in over 200 ms, or appear at once if the wait was under 150 ms or motion is reduced. The frame is drawn only when onboarding was last seen done on the same database, which Electron names by the database file's identity (`window.surfsense.databaseIdentity()`), so a database made since, a first launch and onboarding get the plain background. Docker has no bridge and always gets the plain background. A test keeps the frame's imports off the chat runtime and every feature's `api.ts`, so it stays in the first chunk.
 - Code is grouped by feature under `src/features/`. Each HTTP contract lives in one feature's `api.ts`, next to its hooks and components, and other features import it from there.
 
 ## Codemap
diff --git a/surfsense_local/electron/src/main/index.ts b/surfsense_local/electron/src/main/index.ts
index 195026d2e1..e1182b3281 100644
--- a/surfsense_local/electron/src/main/index.ts
+++ b/surfsense_local/electron/src/main/index.ts
@@ -43,6 +43,7 @@ import { loadLocalePreference } from "./i18n/locale-prefs.ts"
 import { registerLocaleHandlers } from "./i18n/locale-ipc.ts"
 import { registerAboutHandlers } from "./about/about-ipc.ts"
 import { registerDataFolderHandlers } from "./about/data-folder-ipc.ts"
+import { registerStartupFrameHandlers } from "./startup-frame/startup-frame-ipc.ts"
 import { sessionLog } from "../../../sidecars/session-log/session-log.ts"
 import { warnMain } from "../../../sidecars/session-log/main-warn.ts"
 import { registerSessionLogHandlers } from "./session-log/session-log-ipc.ts"
@@ -496,6 +497,11 @@ function main(): void {
         isTrusted: (sender) =>
           mainWindow !== null && sender === mainWindow.webContents,
       })
+      registerStartupFrameHandlers({
+        dataDir: boot.dataDir,
+        isTrusted: (sender) =>
+          mainWindow !== null && sender === mainWindow.webContents,
+      })
       registerMicrophoneHandlers({
         ipc: ipcMain,
         isTrusted: (sender) =>
diff --git a/surfsense_local/electron/src/main/startup-frame/database-identity.test.ts b/surfsense_local/electron/src/main/startup-frame/database-identity.test.ts
new file mode 100644
index 0000000000..6d1b40a10c
--- /dev/null
+++ b/surfsense_local/electron/src/main/startup-frame/database-identity.test.ts
@@ -0,0 +1,42 @@
+import assert from "node:assert/strict"
+import { mkdtempSync, renameSync, rmSync, writeFileSync } from "node:fs"
+import { tmpdir } from "node:os"
+import { join } from "node:path"
+import test from "node:test"
+
+import { databaseIdentity } from "./database-identity.ts"
+
+function dataFolder(): string {
+  return mkdtempSync(join(tmpdir(), "surfsense-identity-"))
+}
+
+test("is absent before the database exists", () => {
+  assert.equal(databaseIdentity(dataFolder()), null)
+})
+
+test("stays the same for the same database", () => {
+  const dir = dataFolder()
+  writeFileSync(join(dir, "surfsense.db"), "")
+  const first = databaseIdentity(dir)
+  assert.equal(typeof first, "string")
+  writeFileSync(join(dir, "surfsense.db"), "written to since")
+  assert.equal(databaseIdentity(dir), first)
+})
+
+test("changes once the database is deleted and made again", () => {
+  const dir = dataFolder()
+  const path = join(dir, "surfsense.db")
+  writeFileSync(path, "")
+  const before = databaseIdentity(dir)
+  // Made before the old one goes, so the filesystem cannot hand back its inode.
+  writeFileSync(`${path}.new`, "")
+  rmSync(path)
+  renameSync(`${path}.new`, path)
+  assert.notEqual(databaseIdentity(dir), before)
+})
+
+test("does not reveal where the data folder is", () => {
+  const dir = dataFolder()
+  writeFileSync(join(dir, "surfsense.db"), "")
+  assert.match(databaseIdentity(dir) ?? "", /^[0-9a-f]{32}$/)
+})
diff --git a/surfsense_local/electron/src/main/startup-frame/database-identity.ts b/surfsense_local/electron/src/main/startup-frame/database-identity.ts
new file mode 100644
index 0000000000..dc03edc590
--- /dev/null
+++ b/surfsense_local/electron/src/main/startup-frame/database-identity.ts
@@ -0,0 +1,20 @@
+import { createHash } from "node:crypto"
+import { statSync } from "node:fs"
+import { join } from "node:path"
+
+/**
+ * Names this data folder's database, so the page can tell a database it has
+ * opened before from one made since: deleted and made again, it is a new one.
+ * Hashed, so the page learns nothing about the disk. Null before it exists.
+ */
+export function databaseIdentity(dataDir: string): string | null {
+  try {
+    const { dev, ino, birthtimeMs } = statSync(join(dataDir, "surfsense.db"))
+    return createHash("sha256")
+      .update(`${dev}:${ino}:${birthtimeMs}`)
+      .digest("hex")
+      .slice(0, 32)
+  } catch {
+    return null
+  }
+}
diff --git a/surfsense_local/electron/src/main/startup-frame/startup-frame-ipc.ts b/surfsense_local/electron/src/main/startup-frame/startup-frame-ipc.ts
new file mode 100644
index 0000000000..58265ee996
--- /dev/null
+++ b/surfsense_local/electron/src/main/startup-frame/startup-frame-ipc.ts
@@ -0,0 +1,16 @@
+import { ipcMain } from "electron"
+
+import { databaseIdentity } from "./database-identity.ts"
+
+// Sync: the page asks while it draws its first frame, before any request returns.
+export function registerStartupFrameHandlers(options: {
+  dataDir: string
+  isTrusted: (sender: Electron.WebContents) => boolean
+}): void {
+  const { dataDir, isTrusted } = options
+  ipcMain.on("startup:database-identity", (event) => {
+    event.returnValue = isTrusted(event.sender)
+      ? databaseIdentity(dataDir)
+      : null
+  })
+}
diff --git a/surfsense_local/electron/src/preload/index.ts b/surfsense_local/electron/src/preload/index.ts
index ff430f83a8..8cfa1f07e2 100644
--- a/surfsense_local/electron/src/preload/index.ts
+++ b/surfsense_local/electron/src/preload/index.ts
@@ -37,6 +37,9 @@ function isSidecarCrash(value: unknown): value is SidecarCrash {
 contextBridge.exposeInMainWorld("surfsense", {
   apiUrl: arg ? arg.slice(FLAG.length) : "http://127.0.0.1:8000",
   platform: process.platform,
+  // Asked when needed: on a first launch the database appears after the page loads.
+  databaseIdentity: (): string | null =>
+    ipcRenderer.sendSync("startup:database-identity") as string | null,
   openDocument: (workspaceId: number, documentId: number): Promise<string> =>
     ipcRenderer.invoke("documents:open", workspaceId, documentId),
   revealDocument: (workspaceId: number, documentId: number): Promise<string> =>
diff --git a/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx b/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
new file mode 100644
index 0000000000..8009211983
--- /dev/null
+++ b/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
@@ -0,0 +1,222 @@
+import { cleanup, screen, within } from "@testing-library/react"
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest"
+
+import { TooltipProvider } from "@/components/ui/tooltip"
+import {
+  expectsDashboard,
+  rememberDashboard,
+} from "@/features/startup/dashboard-hint"
+import { render } from "@/test-utils"
+import { AppBootstrap } from "./app-bootstrap"
+
+const STARTING = { name: "Starting SurfSense" }
+
+function onDatabase(identity: string) {
+  window.surfsense = {
+    apiUrl: "",
+    databaseIdentity: () => identity,
+  } as unknown as Window["surfsense"]
+}
+
+function pendingStartup() {
+  vi.stubGlobal(
+    "fetch",
+    vi.fn(() => new Promise<Response>(() => {}))
+  )
+}
+
+// A completed install with one workspace and no model; `onboardingAfterMs`
+// holds the first answer back, as a slow start does.
+function completedInstall({ onboardingAfterMs = 0 } = {}) {
+  vi.stubGlobal(
+    "fetch",
+    vi.fn(async (input: RequestInfo | URL) => {
+      const path = String(input)
+      if (path === "/llm/onboarding") {
+        await new Promise((resolve) => setTimeout(resolve, onboardingAfterMs))
+        return Response.json({ completed: true })
+      }
+      if (path === "/workspaces") {
+        return Response.json([
+          {
+            id: 1,
+            name: "Research",
+            created_at: "2026-09-09T00:00:00Z",
+            updated_at: "2026-09-09T00:00:00Z",
+          },
+        ])
+      }
+      if (
+        path === "/llm/providers" ||
+        path === "/workspaces/1/chat/threads" ||
+        path === "/workspaces/1/documents?document_type=FILE&document_type=NOTE"
+      ) {
+        return Response.json([])
+      }
+      return Response.json({ detail: "not found" }, { status: 404 })
+    })
+  )
+}
+
+function renderApp() {
+  render(
+    <TooltipProvider>
+      <AppBootstrap />
+    </TooltipProvider>
+  )
+}
+
+beforeEach(() => {
+  vi.stubGlobal(
+    "ResizeObserver",
+    class {
+      observe() {}
+      unobserve() {}
+      disconnect() {}
+    }
+  )
+  // Onboarding's welcome step reads matchMedia.
+  Object.defineProperty(window, "matchMedia", {
+    configurable: true,
+    value: vi.fn().mockReturnValue({
+      matches: false,
+      addEventListener: vi.fn(),
+      removeEventListener: vi.fn(),
+    }),
+  })
+})
+
+afterEach(() => {
+  cleanup()
+  vi.unstubAllGlobals()
+  window.surfsense = undefined
+  localStorage.clear()
+})
+
+describe("startup screen", () => {
+  it("is blank until onboarding has been seen done on this database", () => {
+    onDatabase("a")
+    pendingStartup()
+    renderApp()
+
+    const screenBox = screen.getByRole("status", STARTING)
+    expect(screenBox.textContent).toBe("")
+    expect(within(screenBox).queryByRole("heading")).toBeNull()
+  })
+
+  it("is blank without the desktop, whose data can change under the browser", () => {
+    rememberDashboard()
+    pendingStartup()
+    renderApp()
+
+    expect(screen.getByRole("status", STARTING).textContent).toBe("")
+  })
+
+  it("draws the dashboard's constant parts, and nothing to press, once onboarding is done", () => {
+    onDatabase("a")
+    rememberDashboard()
+    pendingStartup()
+    renderApp()
+
+    const frame = screen.getByRole("status", STARTING)
+    expect(frame.getAttribute("aria-busy")).toBe("true")
+    expect(
+      within(frame).getByRole("heading", { name: "SurfSense" })
+    ).toBeTruthy()
+    expect(within(frame).getByRole("heading", { name: "Studio" })).toBeTruthy()
+    expect(within(frame).queryAllByRole("button")).toEqual([])
+    expect(within(frame).queryByRole("textbox")).toBeNull()
+  })
+
+  it("puts the composer where the dashboard will: centred on a new chat", () => {
+    onDatabase("a")
+    rememberDashboard()
+    pendingStartup()
+    renderApp()
+
+    const frame = screen.getByRole("status", STARTING)
+    expect(frame.querySelector("[data-composer-outline=center]")).toBeTruthy()
+    expect(frame.querySelector("[data-composer-outline=bottom]")).toBeNull()
+  })
+
+  it("puts the composer where the dashboard will: at the bottom of an open thread", () => {
+    onDatabase("a")
+    rememberDashboard()
+    localStorage.setItem("surfsense:last-workspace:v1", "1")
+    localStorage.setItem("surfsense:last-thread:1:v1", "10")
+    pendingStartup()
+    renderApp()
+
+    const frame = screen.getByRole("status", STARTING)
+    expect(frame.querySelector("[data-composer-outline=bottom]")).toBeTruthy()
+    expect(frame.querySelector("[data-composer-outline=center]")).toBeNull()
+  })
+})
+
+describe("startup's memory of onboarding", () => {
+  it("is kept once the API says onboarding is done", async () => {
+    onDatabase("a")
+    completedInstall()
+    renderApp()
+
+    await screen.findByRole(
+      "navigation",
+      { name: "Workspaces" },
+      { timeout: 10_000 }
+    )
+    expect(expectsDashboard()).toBe(true)
+  }, 15_000)
+
+  it("is dropped once the API says onboarding is needed again", async () => {
+    onDatabase("a")
+    rememberDashboard()
+    vi.stubGlobal(
+      "fetch",
+      vi.fn(async (input: RequestInfo | URL) =>
+        String(input) === "/llm/onboarding"
+          ? Response.json({ completed: false })
+          : Response.json({ detail: "not found" }, { status: 404 })
+      )
+    )
+    renderApp()
+
+    await screen.findByRole("button", { name: "Start setting up" })
+    expect(expectsDashboard()).toBe(false)
+  })
+})
+
+describe("startup's reveal", () => {
+  async function workspaceRail() {
+    return screen.findByRole(
+      "navigation",
+      { name: "Workspaces" },
+      { timeout: 10_000 }
+    )
+  }
+
+  it("fades the dashboard's contents into the frame after a wait long enough to see", async () => {
+    onDatabase("a")
+    rememberDashboard()
+    completedInstall({ onboardingAfterMs: 250 })
+    renderApp()
+
+    expect((await workspaceRail()).className).toContain("ss-reveal")
+  }, 15_000)
+
+  it("draws them at once after a wait too short to see", async () => {
+    onDatabase("a")
+    rememberDashboard()
+    completedInstall()
+    renderApp()
+
+    expect((await workspaceRail()).className).not.toContain("ss-reveal")
+  }, 15_000)
+
+  it("never fades without a frame to fade into", async () => {
+    onDatabase("a")
+    completedInstall({ onboardingAfterMs: 250 })
+    renderApp()
+
+    expect((await workspaceRail()).className).not.toContain("ss-reveal")
+  }, 15_000)
+})
diff --git a/surfsense_local/frontend/src/app/app-bootstrap.test.tsx b/surfsense_local/frontend/src/app/app-bootstrap.test.tsx
index 66402b7d1c..9018a54c6f 100644
--- a/surfsense_local/frontend/src/app/app-bootstrap.test.tsx
+++ b/surfsense_local/frontend/src/app/app-bootstrap.test.tsx
@@ -26,24 +26,6 @@ afterEach(() => {
 })
 
 describe("app bootstrap", () => {
-  it("shows the filling logo while startup is pending", () => {
-    vi.stubGlobal(
-      "fetch",
-      vi.fn(() => new Promise<Response>(() => {}))
-    )
-
-    render(
-      <TooltipProvider>
-        <AppBootstrap />
-      </TooltipProvider>
-    )
-
-    // Named for assistive tech; the logo itself is decoration and says nothing.
-    const loader = screen.getByRole("status", { name: "Starting SurfSense" })
-    expect(loader.textContent).toBe("")
-    expect(loader.querySelector(".ss-logo-fill-liquid")).toBeTruthy()
-  })
-
   it("shows onboarding only before it has been completed", async () => {
     vi.stubGlobal(
       "fetch",
diff --git a/surfsense_local/frontend/src/app/app-bootstrap.tsx b/surfsense_local/frontend/src/app/app-bootstrap.tsx
index 31129496e7..e5ff1fa6a8 100644
--- a/surfsense_local/frontend/src/app/app-bootstrap.tsx
+++ b/surfsense_local/frontend/src/app/app-bootstrap.tsx
@@ -17,11 +17,28 @@ import { intl } from "@/i18n/intl"
 
 import { listWorkspaces, type Workspace } from "@/features/workspaces/api"
 
+import {
+  expectsDashboard,
+  forgetDashboard,
+  rememberDashboard,
+} from "@/features/startup/dashboard-hint"
+import { StartupRevealProvider } from "@/features/startup/startup-reveal-provider"
+import { StartupScreen } from "@/features/startup/startup-screen"
+
 import { loadDashboard, type DashboardComponent } from "./load-dashboard"
-import { LogoFillLoader } from "./logo-fill-loader"
 
-type BootstrapState =
-  | { status: "loading" }
+// A wait shorter than this ends before the frame is noticed: the dashboard is
+// drawn at once rather than faded in, which would only make it look slower.
+const REVEAL_AFTER_MS = 150
+
+type Loading = {
+  status: "loading"
+  // Whether the dashboard's frame is drawn, rather than the plain background.
+  frame: boolean
+  since: number
+}
+
+type Fetched =
   | { status: "onboarding-required" }
   | {
       status: "ready"
@@ -35,6 +52,25 @@ type BootstrapState =
     }
   | { status: "error"; message: string }
 
+type BootstrapState =
+  | Loading
+  | Exclude<Fetched, { status: "ready" }>
+  | (Extract<Fetched, { status: "ready" }> & { reveal: boolean })
+
+function startLoading(): Loading {
+  return { status: "loading", frame: expectsDashboard(), since: Date.now() }
+}
+
+// Fades the dashboard into a frame that was on screen long enough to see.
+function settle(current: BootstrapState, next: Fetched): BootstrapState {
+  if (next.status !== "ready") return next
+  const reveal =
+    current.status === "loading" &&
+    current.frame &&
+    Date.now() - current.since >= REVEAL_AFTER_MS
+  return { ...next, reveal }
+}
+
 function messageFrom(error: unknown) {
   return error instanceof Error
     ? error.message
@@ -44,7 +80,7 @@ function messageFrom(error: unknown) {
       })
 }
 
-async function fetchBootstrapState(): Promise<BootstrapState> {
+async function fetchBootstrapState(): Promise<Fetched> {
   // Fetched alongside the data rather than after it: one wait, one loader.
   const dashboard = loadDashboard()
   // Handled where it is awaited; this only keeps an early exit from leaving
@@ -53,8 +89,10 @@ async function fetchBootstrapState(): Promise<BootstrapState> {
   try {
     const onboarding = await getOnboardingStatus()
     if (!onboarding.completed) {
+      forgetDashboard()
       return { status: "onboarding-required" }
     }
+    rememberDashboard()
     const [selection, workspaces] = await Promise.all([
       getGenerationSelection(),
       listWorkspaces(),
@@ -76,28 +114,13 @@ async function fetchBootstrapState(): Promise<BootstrapState> {
   }
 }
 
-function GlobalLoader() {
-  return (
-    <main
-      className="flex h-full items-center justify-center bg-app-shell select-none"
-      role="status"
-      aria-label={intl.formatMessage({
-        id: "app_bootstrap_loader_aria",
-        defaultMessage: "Starting SurfSense",
-      })}
-    >
-      <LogoFillLoader />
-    </main>
-  )
-}
-
 export function AppBootstrap() {
-  const [state, setState] = useState<BootstrapState>({ status: "loading" })
+  const [state, setState] = useState<BootstrapState>(startLoading)
 
   useEffect(() => {
     let active = true
     void fetchBootstrapState().then((next) => {
-      if (active) setState(next)
+      if (active) setState((current) => settle(current, next))
     })
     return () => {
       active = false
@@ -105,23 +128,26 @@ export function AppBootstrap() {
   }, [])
 
   if (state.status === "loading") {
-    return <GlobalLoader />
+    return <StartupScreen frame={state.frame} />
   }
 
   if (state.status === "onboarding-required") {
     return (
       <OnboardingPage
         onComplete={(selection) => {
-          setState({ status: "loading" })
+          rememberDashboard()
+          setState(startLoading())
           void Promise.all([listWorkspaces(), loadDashboard()])
             .then(([workspaces, Dashboard]) =>
-              setState({
-                status: "ready",
-                Dashboard,
-                selection,
-                availability: { status: "available" },
-                workspaces,
-              })
+              setState((current) =>
+                settle(current, {
+                  status: "ready",
+                  Dashboard,
+                  selection,
+                  availability: { status: "available" },
+                  workspaces,
+                })
+              )
             )
             .catch((error: unknown) =>
               setState({ status: "error", message: messageFrom(error) })
@@ -148,8 +174,10 @@ export function AppBootstrap() {
               className="mt-3"
               variant="outline"
               onClick={() => {
-                setState({ status: "loading" })
-                void fetchBootstrapState().then(setState)
+                setState(startLoading())
+                void fetchBootstrapState().then((next) =>
+                  setState((current) => settle(current, next))
+                )
               }}
             >
               {intl.formatMessage({
@@ -167,20 +195,24 @@ export function AppBootstrap() {
   // loader to flash between startup and the first screen.
   const { Dashboard } = state
   return (
-    <Dashboard
-      selection={state.selection}
-      initialAvailability={state.availability}
-      initialWorkspaces={state.workspaces}
-      onModelUnavailable={() =>
-        setState((current) =>
-          current.status === "ready" ? { ...current, selection: null } : current
-        )
-      }
-      onModelSelected={(selection) =>
-        setState((current) =>
-          current.status === "ready" ? { ...current, selection } : current
-        )
-      }
-    />
+    <StartupRevealProvider reveal={state.reveal}>
+      <Dashboard
+        selection={state.selection}
+        initialAvailability={state.availability}
+        initialWorkspaces={state.workspaces}
+        onModelUnavailable={() =>
+          setState((current) =>
+            current.status === "ready"
+              ? { ...current, selection: null }
+              : current
+          )
+        }
+        onModelSelected={(selection) =>
+          setState((current) =>
+            current.status === "ready" ? { ...current, selection } : current
+          )
+        }
+      />
+    </StartupRevealProvider>
   )
 }
diff --git a/surfsense_local/frontend/src/app/logo-fill-loader.tsx b/surfsense_local/frontend/src/app/logo-fill-loader.tsx
deleted file mode 100644
index 870748aee6..0000000000
--- a/surfsense_local/frontend/src/app/logo-fill-loader.tsx
+++ /dev/null
@@ -1,55 +0,0 @@
-import { useState, type CSSProperties } from "react"
-
-import surfSenseLogo from "@/surfsense-logo.svg"
-import { cn } from "@/lib/utils"
-
-/** One full tide, as `.ss-logo-fill-liquid` animates it in `index.css`. */
-const FILL_CYCLE_MS = 2200
-
-// The logo is the mask, as in onboarding's brand row, so its paths are drawn
-// once by the browser rather than copied into the page.
-const LOGO_MASK = {
-  maskImage: `url(${surfSenseLogo})`,
-  maskPosition: "center",
-  maskRepeat: "no-repeat",
-  maskSize: "contain",
-} as const
-
-/**
- * The logo filling with a rippling tide, looped until the app is ready: a
- * faint logo, and a solid one rising through it behind a wave. The motion is
- * in `index.css` (`ss-logo-fill-*`); under reduced motion it rests full.
- */
-export function LogoFillLoader({ className }: { className?: string }) {
-  // Where the tide is on the page clock, read once on mount: every loader
-  // agrees on the phase, so a remount continues rather than restarts.
-  const [offset] = useState(() => -(performance.now() % FILL_CYCLE_MS))
-
-  return (
-    <div
-      aria-hidden="true"
-      className={cn("relative size-14 overflow-hidden", className)}
-      style={
-        {
-          ...LOGO_MASK,
-          "--ss-logo-fill-offset": `${offset}ms`,
-        } as CSSProperties
-      }
-    >
-      <div className="absolute inset-0 bg-foreground/15" />
-      <div className="ss-logo-fill-liquid absolute inset-x-0 top-0 flex h-[calc(100%+0.5rem)] flex-col">
-        {/* Two periods of the wave, so sliding it by half loops seamlessly. */}
-        <svg
-          aria-hidden="true"
-          focusable="false"
-          viewBox="0 0 200 16"
-          preserveAspectRatio="none"
-          className="ss-logo-fill-wave block h-2 w-[200%] shrink-0 fill-foreground"
-        >
-          <path d="M0 8 Q25 0 50 8 T100 8 T150 8 T200 8 V16 H0 Z" />
-        </svg>
-        <div className="flex-1 bg-foreground" />
-      </div>
-    </div>
-  )
-}
diff --git a/surfsense_local/frontend/src/features/chat/chat-composer.tsx b/surfsense_local/frontend/src/features/chat/chat-composer.tsx
index 39018a7d9f..1c0df78dbd 100644
--- a/surfsense_local/frontend/src/features/chat/chat-composer.tsx
+++ b/surfsense_local/frontend/src/features/chat/chat-composer.tsx
@@ -14,9 +14,17 @@ import { PluginMentions } from "@/features/plugins/in-thread/plugin-mentions"
 import { cn } from "@/lib/utils"
 import { intl } from "@/i18n/intl"
 import type { PluginsStart } from "@/features/plugins/plugins-settings"
+import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 
 import { ComposerImage } from "./attached-image"
 import { ComposerAddMenu } from "./composer-add-menu"
+import {
+  COMPOSER_CLASS,
+  COMPOSER_CONTROLS_CLASS,
+  COMPOSER_FOOTER_CLASS,
+  composerBoxClass,
+  composerInputClass,
+} from "./composer-layout"
 import { ModePicker } from "./modes/mode-picker"
 import { useNewChatMode } from "./modes/new-chat-mode"
 import { ModelPicker, modelControlButtonClassName } from "./model-picker"
@@ -209,6 +217,7 @@ export function ChatComposer({
   const atLimit = useAuiState(
     ({ composer }) => composer.text.length >= QUESTION_MAX_CHARS
   )
+  const reveal = useStartupReveal()
   const dictationKeys = useDictationKeys(dictation)
   // While the microphone starts, the mic's spinner says so and nothing else moves.
   const recording = useDictationStarted(dictation)
@@ -253,10 +262,7 @@ export function ChatComposer({
     />
   )
   return (
-    <div
-      className="relative mx-auto w-full max-w-2xl"
-      data-composer-placement={placement}
-    >
+    <div className={COMPOSER_CLASS} data-composer-placement={placement}>
       {notice ? (
         // Tucked behind the composer, which paints over its lower edge: out of
         // flow, so the composer keeps its place and its own shape.
@@ -271,10 +277,9 @@ export function ChatComposer({
       ) : null}
       <ComposerPrimitive.Root
         className={cn(
-          // No top padding: the text scrolls up to the border, not to a strip
-          // above it, and holds that space in its own padding instead.
-          "relative rounded-2xl border bg-card px-1.5 pb-1.5 shadow-sm transition-colors focus-within:border-ring/40 hover:border-ring/40",
-          placement === "bottom" && "flex items-end gap-2"
+          composerBoxClass(placement),
+          "transition-colors focus-within:border-ring/40 hover:border-ring/40",
+          reveal
         )}
       >
         <ComposerPrimitive.Unstable_TriggerPopoverRoot>
@@ -292,12 +297,7 @@ export function ChatComposer({
             readOnly={dictationKeys.recording}
             cancelOnEscape={!dictationKeys.recording}
             onKeyDown={dictationKeys.onKeyDown}
-            className={cn(
-              "max-h-44 resize-none bg-transparent px-2 pt-4 pb-2.5 text-sm outline-none placeholder:text-muted-foreground",
-              placement === "center"
-                ? "block min-h-12 w-full"
-                : "min-h-10 flex-1"
-            )}
+            className={composerInputClass(placement)}
             placeholder={
               blockedPlaceholder !== undefined
                 ? blockedPlaceholder
@@ -328,7 +328,7 @@ export function ChatComposer({
           {placement === "center" ? (
             // Its own row under the text, never over it: long text scrolls
             // inside the box and stops above the controls.
-            <div className="flex items-center justify-between gap-2">
+            <div className={COMPOSER_CONTROLS_CLASS}>
               <div className="flex shrink-0 items-center gap-1">
                 {addMenu("")}
                 {model ? <ModePicker model={model} threadMode={null} /> : null}
@@ -381,7 +381,7 @@ export function ChatComposer({
         </p>
       ) : null}
       {placement === "bottom" ? (
-        <div className="mt-1 flex min-h-7 items-center justify-between gap-3 px-2">
+        <div className={cn(COMPOSER_FOOTER_CLASS, reveal)}>
           {/* Keeps its line and leaves the mode and the model name what is
           left, never less than 13rem, so a long translation truncates the
           name before wrapping. */}
diff --git a/surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts b/surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts
new file mode 100644
index 0000000000..8c49f79ba5
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts
@@ -0,0 +1,19 @@
+// The conversation's scroll area and its sticky composer footer, shared by
+// ChatViewport and the startup frame, which draws them without the runtime.
+
+export const THREAD_BODY_CLASS = "relative flex min-h-0 flex-1 flex-col"
+
+export const VIEWPORT_BOX_CLASS =
+  "relative flex min-h-0 w-full min-w-0 flex-1 flex-col"
+
+// Top edge only: a bottom mask would fade the sticky composer.
+export const VIEWPORT_CLASS =
+  "flex min-h-0 w-full min-w-0 flex-1 scroll-fade-t flex-col overflow-y-auto px-4 [--scroll-fade-reveal:24px] scroll-fade-6"
+
+// The gutter moves the centred column on systems with classic scrollbars.
+export const VIEWPORT_STYLE = { scrollbarGutter: "stable" } as const
+
+export const VIEWPORT_FOOTER_CLASS =
+  "sticky bottom-0 z-20 -mx-4 mt-auto shrink-0 bg-gradient-to-t from-background via-background to-transparent px-4 pb-1"
+
+export const VIEWPORT_FOOTER_INNER_CLASS = "relative mx-auto w-full max-w-3xl"
diff --git a/surfsense_local/frontend/src/features/chat/chat-viewport.tsx b/surfsense_local/frontend/src/features/chat/chat-viewport.tsx
index a45017d08c..a91a3f6c2d 100644
--- a/surfsense_local/frontend/src/features/chat/chat-viewport.tsx
+++ b/surfsense_local/frontend/src/features/chat/chat-viewport.tsx
@@ -5,6 +5,14 @@ import { ArrowDownIcon } from "@/components/ui/icons"
 import { intl } from "@/i18n/intl"
 import { cn } from "@/lib/utils"
 
+import {
+  VIEWPORT_BOX_CLASS,
+  VIEWPORT_CLASS,
+  VIEWPORT_FOOTER_CLASS,
+  VIEWPORT_FOOTER_INNER_CLASS,
+  VIEWPORT_STYLE,
+} from "./chat-viewport-layout"
+
 // Raised clear of a notice tucked behind the composer's top edge, whose visible
 // part is about 28px tall.
 function ScrollToBottom({ raised }: { raised: boolean }) {
@@ -39,12 +47,11 @@ export function ChatViewport({
   footerHasNotice?: boolean
 }) {
   return (
-    <div className="relative flex min-h-0 w-full min-w-0 flex-1 flex-col">
+    <div className={VIEWPORT_BOX_CLASS}>
       <ThreadPrimitive.Viewport
         turnAnchor="top"
-        // Top edge only: a bottom mask would fade the sticky composer.
-        className="flex min-h-0 w-full min-w-0 flex-1 scroll-fade-t flex-col overflow-y-auto px-4 [--scroll-fade-reveal:24px] scroll-fade-6"
-        style={{ scrollbarGutter: "stable" }}
+        className={VIEWPORT_CLASS}
+        style={VIEWPORT_STYLE}
         autoScroll
         scrollToBottomOnRunStart
         scrollToBottomOnInitialize
@@ -53,8 +60,8 @@ export function ChatViewport({
       >
         {children}
         {footer ? (
-          <ThreadPrimitive.ViewportFooter className="sticky bottom-0 z-20 -mx-4 mt-auto shrink-0 bg-gradient-to-t from-background via-background to-transparent px-4 pb-1">
-            <div className="relative mx-auto w-full max-w-3xl">
+          <ThreadPrimitive.ViewportFooter className={VIEWPORT_FOOTER_CLASS}>
+            <div className={VIEWPORT_FOOTER_INNER_CLASS}>
               <ScrollToBottom raised={footerHasNotice} />
               {footer}
             </div>
diff --git a/surfsense_local/frontend/src/features/chat/composer-layout.ts b/surfsense_local/frontend/src/features/chat/composer-layout.ts
new file mode 100644
index 0000000000..f6faf262ce
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/composer-layout.ts
@@ -0,0 +1,33 @@
+import { cn } from "@/lib/utils"
+
+// The composer's geometry, shared by ChatComposer and the startup frame's
+// outline of it, so the box the frame draws is the box that replaces it.
+
+export type ComposerPlacement = "center" | "bottom"
+
+export const COMPOSER_CLASS = "relative mx-auto w-full max-w-2xl"
+
+// No top padding: the text scrolls up to the border, not to a strip above
+// it, and holds that space in its own padding instead.
+export function composerBoxClass(placement: ComposerPlacement) {
+  return cn(
+    "relative rounded-2xl border bg-card px-1.5 pb-1.5 shadow-sm",
+    placement === "bottom" && "flex items-end gap-2"
+  )
+}
+
+// One line tall at rest: an outline holds a line of nothing to match it.
+export function composerInputClass(placement: ComposerPlacement) {
+  return cn(
+    "max-h-44 resize-none bg-transparent px-2 pt-4 pb-2.5 text-sm outline-none placeholder:text-muted-foreground",
+    placement === "center" ? "block min-h-12 w-full" : "min-h-10 flex-1"
+  )
+}
+
+// At least as tall as its icon-lg buttons, so the outline, which has none,
+// keeps the height they give it.
+export const COMPOSER_CONTROLS_CLASS =
+  "flex min-h-9 items-center justify-between gap-2"
+
+export const COMPOSER_FOOTER_CLASS =
+  "mt-1 flex min-h-7 items-center justify-between gap-3 px-2"
diff --git a/surfsense_local/frontend/src/features/chat/composer-outline.tsx b/surfsense_local/frontend/src/features/chat/composer-outline.tsx
new file mode 100644
index 0000000000..68c92eca96
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/composer-outline.tsx
@@ -0,0 +1,36 @@
+import {
+  COMPOSER_CLASS,
+  COMPOSER_CONTROLS_CLASS,
+  COMPOSER_FOOTER_CLASS,
+  composerBoxClass,
+  composerInputClass,
+  type ComposerPlacement,
+} from "./composer-layout"
+
+/**
+ * The composer's empty box, the size ChatComposer draws it, for the startup
+ * frame: its text and controls fade in when the real one takes its place.
+ */
+export function ComposerOutline({
+  placement,
+}: {
+  placement: ComposerPlacement
+}) {
+  return (
+    <div
+      aria-hidden="true"
+      className={COMPOSER_CLASS}
+      data-composer-outline={placement}
+    >
+      <div className={composerBoxClass(placement)}>
+        <div className={composerInputClass(placement)}>&nbsp;</div>
+        {placement === "center" ? (
+          <div className={COMPOSER_CONTROLS_CLASS} />
+        ) : null}
+      </div>
+      {placement === "bottom" ? (
+        <div className={COMPOSER_FOOTER_CLASS} />
+      ) : null}
+    </div>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/chat/conversation-column.tsx b/surfsense_local/frontend/src/features/chat/conversation-column.tsx
new file mode 100644
index 0000000000..4693694d92
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/conversation-column.tsx
@@ -0,0 +1,39 @@
+import type { ReactNode } from "react"
+
+import { ColumnHeader } from "@/components/ui/column-header"
+import { intl } from "@/i18n/intl"
+import { cn } from "@/lib/utils"
+
+/**
+ * The conversation's box and its header line, which the startup frame draws
+ * empty. `heading` precedes the header; `headerClassName` styles its line.
+ */
+export function ConversationColumn({
+  heading,
+  header,
+  headerClassName,
+  children,
+}: {
+  heading?: ReactNode
+  header?: ReactNode
+  headerClassName?: string
+  children: ReactNode
+}) {
+  return (
+    <section
+      className="flex h-full min-w-0 flex-col bg-background"
+      aria-label={intl.formatMessage({
+        id: "chat_thread_panel_conversation_aria",
+        defaultMessage: "Conversation",
+      })}
+    >
+      {heading}
+      <header>
+        <ColumnHeader className={cn("px-5", headerClassName)}>
+          {header}
+        </ColumnHeader>
+      </header>
+      {children}
+    </section>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/chat/last-thread.ts b/surfsense_local/frontend/src/features/chat/last-thread.ts
new file mode 100644
index 0000000000..47d021e463
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/last-thread.ts
@@ -0,0 +1,27 @@
+function lastThreadKey(workspaceId: number) {
+  return `surfsense:last-thread:${workspaceId}:v1`
+}
+
+const NEW_CHAT = "new"
+
+/** The thread open last in a workspace; null for a new chat. */
+export function readLastThread(workspaceId: number): number | null {
+  try {
+    const id = Number(localStorage.getItem(lastThreadKey(workspaceId)))
+    if (Number.isInteger(id) && id > 0) return id
+  } catch {
+    // Private browsing: treat as a new chat.
+  }
+  return null
+}
+
+export function rememberThread(workspaceId: number, threadId: number | null) {
+  try {
+    localStorage.setItem(
+      lastThreadKey(workspaceId),
+      threadId === null ? NEW_CHAT : String(threadId)
+    )
+  } catch {
+    // Selection remains valid for this session when storage is unavailable.
+  }
+}
diff --git a/surfsense_local/frontend/src/features/chat/messages-skeleton.tsx b/surfsense_local/frontend/src/features/chat/messages-skeleton.tsx
new file mode 100644
index 0000000000..3bc0f85935
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/messages-skeleton.tsx
@@ -0,0 +1,17 @@
+import { Skeleton } from "@/components/ui/skeleton"
+
+/** A thread's messages while they load: one question and the start of a reply. */
+export function MessagesSkeleton() {
+  return (
+    <div className="mx-auto flex w-full max-w-2xl flex-col">
+      <div className="flex flex-col items-end px-6 py-3">
+        <Skeleton className="h-10 w-[42%] rounded-2xl rounded-br-md" />
+      </div>
+      <div className="flex flex-col items-start gap-2 px-6 py-4">
+        <Skeleton className="h-4 w-[92%]" />
+        <Skeleton className="h-4 w-[80%]" />
+        <Skeleton className="h-4 w-[58%]" />
+      </div>
+    </div>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/chat/thread-panel.tsx b/surfsense_local/frontend/src/features/chat/thread-panel.tsx
index e9aeb26c28..073027205f 100644
--- a/surfsense_local/frontend/src/features/chat/thread-panel.tsx
+++ b/surfsense_local/frontend/src/features/chat/thread-panel.tsx
@@ -22,22 +22,25 @@ import {
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu"
 import { Input } from "@/components/ui/input"
-import { Skeleton } from "@/components/ui/skeleton"
 import { TypewriterText } from "@/components/typewriter-text"
 import type { Dictation } from "@/features/dictation/dictation-adapter"
 import type { ModelSelection } from "@/features/models/selection/api"
 import { useStableCallback } from "@/hooks/use-stable-callback"
-import { ColumnHeader } from "@/components/ui/column-header"
 import { intl } from "@/i18n/intl"
 import type { PluginsStart } from "@/features/plugins/plugins-settings"
+import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 import type { ChatThread } from "./api"
 import { ChatComposer } from "./chat-composer"
 import { ChatViewport } from "./chat-viewport"
+import { THREAD_BODY_CLASS } from "./chat-viewport-layout"
+import { ConversationColumn } from "./conversation-column"
 import { composerAttachmentOf } from "./image-attachments"
 import { LiveThreadRuntime, type LiveThreadSource } from "./live-thread-runtime"
 import { AssistantMessage, UserMessage } from "./message"
+import { MessagesSkeleton } from "./messages-skeleton"
 import type { Citation } from "./sse"
 import type { ConversationView } from "./use-chat-runtime"
+import { ThreadWelcome } from "./thread-welcome"
 
 // Shared, so a reply without citations passes the same props at every render.
 const NO_CITATIONS: Citation[] = []
@@ -105,17 +108,6 @@ function useMessageRenderer({
   )
 }
 
-function ThreadWelcome({ composer }: { composer: ReactNode }) {
-  return (
-    <div className="flex min-h-0 flex-1">
-      <div className="mx-auto grid h-full w-full max-w-3xl grid-rows-[minmax(0,1fr)_auto_minmax(0,1fr)]">
-        <div />
-        {composer}
-      </div>
-    </div>
-  )
-}
-
 function ComposerDraftLifecycle({
   view,
   dictation,
@@ -215,6 +207,7 @@ export function ThreadPanel({
   // At the header's end: the right panel's toggle while the panel is hidden.
   headerEnd?: ReactNode
 }) {
+  const reveal = useStartupReveal()
   const titleInputRef = useRef<HTMLInputElement>(null)
   const headingId = useId()
   const ignoreMenuFocusRef = useRef(false)
@@ -297,23 +290,20 @@ export function ThreadPanel({
   return (
     <LiveThreadRuntime {...live}>
       <ComposerDraftLifecycle view={view} dictation={dictation} />
-      <section
-        className="flex h-full min-w-0 flex-col bg-background"
-        aria-label={intl.formatMessage({
-          id: "chat_thread_panel_conversation_aria",
-          defaultMessage: "Conversation",
-        })}
-      >
-        {/* The title control is a button that renames, so the heading is its
-            own element. Focus stays on the composer, which it describes. A
-            new chat has no conversation to name yet. */}
-        {thread == null ? null : (
-          <h2 id={headingId} className="sr-only">
-            {title}
-          </h2>
-        )}
-        <header>
-          <ColumnHeader className="px-5">
+      <ConversationColumn
+        // The title control is a button that renames, so the heading is its
+        // own element. Focus stays on the composer, which it describes. A
+        // new chat has no conversation to name yet.
+        heading={
+          thread == null ? null : (
+            <h2 id={headingId} className="sr-only">
+              {title}
+            </h2>
+          )
+        }
+        headerClassName={reveal}
+        header={
+          <>
             {thread == null ? null : editing ? (
               <Input
                 ref={titleInputRef}
@@ -427,23 +417,12 @@ export function ThreadPanel({
             {headerEnd ? (
               <div className="-mr-2 ml-auto pl-3">{headerEnd}</div>
             ) : null}
-          </ColumnHeader>
-        </header>
-
-        <ThreadPrimitive.Root className="relative flex min-h-0 flex-1 flex-col">
+          </>
+        }
+      >
+        <ThreadPrimitive.Root className={THREAD_BODY_CLASS}>
           <ChatViewport footer={bottomFooter} footerHasNotice={notice != null}>
-            {isLoading ? (
-              <div className="mx-auto flex w-full max-w-2xl flex-col">
-                <div className="flex flex-col items-end px-6 py-3">
-                  <Skeleton className="h-10 w-[42%] rounded-2xl rounded-br-md" />
-                </div>
-                <div className="flex flex-col items-start gap-2 px-6 py-4">
-                  <Skeleton className="h-4 w-[92%]" />
-                  <Skeleton className="h-4 w-[80%]" />
-                  <Skeleton className="h-4 w-[58%]" />
-                </div>
-              </div>
-            ) : null}
+            {isLoading ? <MessagesSkeleton /> : null}
 
             {view.status === "new" ? (
               <ThreadWelcome composer={composer("center")} />
@@ -456,7 +435,7 @@ export function ThreadPanel({
             ) : null}
           </ChatViewport>
         </ThreadPrimitive.Root>
-      </section>
+      </ConversationColumn>
     </LiveThreadRuntime>
   )
 }
diff --git a/surfsense_local/frontend/src/features/chat/thread-welcome.tsx b/surfsense_local/frontend/src/features/chat/thread-welcome.tsx
new file mode 100644
index 0000000000..538d67f535
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/thread-welcome.tsx
@@ -0,0 +1,13 @@
+import type { ReactNode } from "react"
+
+/** A new chat: the composer alone, centred in the conversation. */
+export function ThreadWelcome({ composer }: { composer: ReactNode }) {
+  return (
+    <div className="flex min-h-0 flex-1">
+      <div className="mx-auto grid h-full w-full max-w-3xl grid-rows-[minmax(0,1fr)_auto_minmax(0,1fr)]">
+        <div />
+        {composer}
+      </div>
+    </div>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/chat/use-chat-runtime.ts b/surfsense_local/frontend/src/features/chat/use-chat-runtime.ts
index bd16d5ae85..77e75a47d6 100644
--- a/surfsense_local/frontend/src/features/chat/use-chat-runtime.ts
+++ b/surfsense_local/frontend/src/features/chat/use-chat-runtime.ts
@@ -58,6 +58,7 @@ import {
   type RunState,
 } from "./runs/run-store"
 import { storedUploads } from "./runs/stored-uploads"
+import { readLastThread, rememberThread } from "./last-thread"
 import { markRead, markUnread, readUnread } from "./runs/unread-replies"
 import type { ChatErrorKind, ChatStreamEvent } from "./sse"
 import { composedThread, isOptimistic, OPTIMISTIC_ID } from "./thread-messages"
@@ -129,12 +130,6 @@ function listedRunState(thread: ChatThread): RunState | null {
   return { state: "running" }
 }
 
-function lastThreadKey(workspaceId: number) {
-  return `surfsense:last-thread:${workspaceId}:v1`
-}
-
-const NEW_CHAT = "new"
-
 export type ConversationView =
   // `returned` is what a refused send hands back to the composer.
   | { status: "new"; returned?: ReturnedDraft }
@@ -143,29 +138,9 @@ export type ConversationView =
 
 export type ReturnedDraft = { text: string; images: ImageUpload[] }
 
-function rememberThread(workspaceId: number, threadId: number | null) {
-  try {
-    localStorage.setItem(
-      lastThreadKey(workspaceId),
-      threadId === null ? NEW_CHAT : String(threadId)
-    )
-  } catch {
-    // Selection remains valid for this session when storage is unavailable.
-  }
-}
-
 function readStoredView(workspaceId: number): ConversationView {
-  try {
-    const stored = localStorage.getItem(lastThreadKey(workspaceId))
-    if (stored === NEW_CHAT) return { status: "new" }
-    const id = Number(stored)
-    if (Number.isInteger(id) && id > 0) {
-      return { status: "active", threadId: id }
-    }
-  } catch {
-    // Private browsing: treat as a new chat.
-  }
-  return { status: "new" }
+  const threadId = readLastThread(workspaceId)
+  return threadId === null ? { status: "new" } : { status: "active", threadId }
 }
 
 function stoppedBeforeAWord(pair: LivePair) {
diff --git a/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx b/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx
index 615c665a81..90de0ee814 100644
--- a/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx
@@ -65,7 +65,7 @@ export function SidebarColumn({
     <div
       id={id}
       className={cn(
-        "flex h-full min-h-0 min-w-68 flex-col transition-[width] duration-[240ms] ease-[cubic-bezier(0.4,0,0.2,1)] motion-reduce:transition-none",
+        "flex h-full min-h-0 min-w-68 flex-col transition-[width] duration-240 ease-in-out motion-reduce:transition-none",
         !columns.animate && "transition-none"
       )}
       style={{ width: columns.sidebar.width }}
@@ -75,6 +75,13 @@ export function SidebarColumn({
   )
 }
 
+/** The chat's column, which takes what the other two leave. */
+export function ChatColumn({ children }: { children: ReactNode }) {
+  return (
+    <div className="flex min-h-0 min-w-130 flex-1 flex-col">{children}</div>
+  )
+}
+
 /** The draggable edge of the sidebar, or of the right panel. */
 export function ColumnEdge({
   column,
diff --git a/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx b/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx
index b38cc57f49..e483c35b44 100644
--- a/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx
@@ -58,8 +58,10 @@ import { useStudio } from "@/features/studio/use-studio"
 import { TitleBarUpdate } from "@/features/updates/title-bar-update"
 import type { Workspace } from "@/features/workspaces/api"
 import { useWorkspaces } from "@/features/workspaces/use-workspaces"
+import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 import { useStableCallback } from "@/hooks/use-stable-callback"
 import { intl } from "@/i18n/intl"
+import { cn } from "@/lib/utils"
 import { WorkspaceRail } from "@/features/workspaces/workspace-rail"
 import type { PluginsStart } from "@/features/plugins/plugins-settings"
 import {
@@ -69,6 +71,7 @@ import {
   writeSourcePreview,
 } from "./chrome-prefs"
 import {
+  ChatColumn,
   ColumnEdge,
   DashboardColumns,
   RightPanelRail,
@@ -77,6 +80,7 @@ import {
 import { LeftSidebar, type SidebarNavAction } from "./left-sidebar"
 import { RightPanel } from "./right-panel"
 import { RightPanelToggle } from "./right-panel-toggle"
+import { DashboardShell } from "./dashboard-shell"
 import { SidebarFooter } from "./sidebar-footer"
 
 // Clicking "N sources" in the composer used to switch the right rail to its
@@ -128,6 +132,7 @@ function WorkspaceDashboard({
   onOpenPlugins: (start?: PluginsStart) => void
   modelsVisited: number
 }) {
+  const reveal = useStartupReveal()
   const [inspect, setInspect] = useState<Inspect>(null)
   const [rightPanelOpen, setRightPanelOpen] = useState(readRightPanelOpen)
   const [sourcePreviewId, setSourcePreviewId] = useState<number | null>(() =>
@@ -390,7 +395,7 @@ function WorkspaceDashboard({
     <OfficeOfferContext.Provider value={officeOffer}>
       <OpenArtifactContext.Provider value={openArtifact}>
         <div className="titlebar-controls">
-          <div className="titlebar-controls-end">
+          <div className={cn("titlebar-controls-end", reveal)}>
             <TitleBarUpdate />
             {/* Docker's browser has no title bar: there the toggle sits on a heading's line. */}
             {IN_CONTAINER
@@ -450,7 +455,7 @@ function WorkspaceDashboard({
             })}
             controls={LEFT_COLUMN_ID}
           />
-          <div className="flex min-h-0 min-w-130 flex-1 flex-col">
+          <ChatColumn>
             <AgentApprovalsProvider
               approvals={chat.approvals}
               answer={chat.answerApproval}
@@ -502,7 +507,7 @@ function WorkspaceDashboard({
                 }}
               />
             </AgentApprovalsProvider>
-          </div>
+          </ChatColumn>
           {/* Collapsed, the rail has no edge to drag. */}
           {rightPanelOpen ? (
             <ColumnEdge
@@ -764,7 +769,7 @@ export function DashboardPage({
   return (
     // Above the workspace, so a mode picked for the next chat holds across them.
     <NewChatModeProvider>
-      <main className="relative grid h-full min-w-292 grid-cols-[56px_minmax(0,1fr)] overflow-hidden bg-app-shell">
+      <DashboardShell>
         <WorkspaceRail
           workspaces={workspaces.workspaces}
           activeWorkspaceId={workspaces.activeWorkspace.id}
@@ -850,7 +855,7 @@ export function DashboardPage({
             </Button>
           </Alert>
         ) : null}
-      </main>
+      </DashboardShell>
     </NewChatModeProvider>
   )
 }
diff --git a/surfsense_local/frontend/src/features/dashboard/dashboard-shell.tsx b/surfsense_local/frontend/src/features/dashboard/dashboard-shell.tsx
new file mode 100644
index 0000000000..05cee26c1c
--- /dev/null
+++ b/surfsense_local/frontend/src/features/dashboard/dashboard-shell.tsx
@@ -0,0 +1,22 @@
+import type { ComponentProps } from "react"
+
+import { cn } from "@/lib/utils"
+
+/**
+ * The window's grid: the workspace rail, then the columns. Shared with the
+ * startup frame, so the dashboard's first frame lands on the frame's lines.
+ */
+export function DashboardShell({
+  className,
+  ...props
+}: ComponentProps<"main">) {
+  return (
+    <main
+      className={cn(
+        "relative grid h-full min-w-292 grid-cols-[56px_minmax(0,1fr)] overflow-hidden bg-app-shell",
+        className
+      )}
+      {...props}
+    />
+  )
+}
diff --git a/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx b/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
index f6793b96d7..440bdb0744 100644
--- a/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
@@ -11,8 +11,11 @@ import { Button } from "@/components/ui/button"
 import { ChatsDialog } from "@/features/chat/chats-dialog"
 import type { ChatThread } from "@/features/chat/api"
 import type { RunState } from "@/features/chat/runs/run-store"
-import { ColumnHeader } from "@/components/ui/column-header"
+import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 import { intl } from "@/i18n/intl"
+import { cn } from "@/lib/utils"
+
+import { SidebarFrame } from "./sidebar-frame"
 
 // A row rendered below "New chat" with the same look. Add an entry here (or
 // pass one through `actions`) rather than hand-rolling another Button.
@@ -168,16 +171,12 @@ export const LeftSidebar = memo(function LeftSidebar({
   footer?: ReactNode
 }) {
   const [chatsOpen, setChatsOpen] = useState(false)
+  const reveal = useStartupReveal()
 
   return (
-    <aside className="flex h-full min-w-0 flex-col border-r bg-background select-none">
-      <header>
-        <ColumnHeader>
-          <h2 className="truncate px-1 brand-wordmark text-xl text-foreground select-none">
-            SurfSense
-          </h2>
-        </ColumnHeader>
-        <div className="flex flex-col px-3 pt-2">
+    <SidebarFrame
+      nav={
+        <div className={cn("flex flex-col px-3 pt-2", reveal)}>
           <SidebarNavButton
             label={intl.formatMessage({
               id: "dashboard_sidebar_new_chat_button",
@@ -199,11 +198,17 @@ export const LeftSidebar = memo(function LeftSidebar({
             <SidebarNavButton key={key} {...action} />
           ))}
         </div>
-      </header>
-      <div className="flex min-h-0 min-w-0 flex-1 flex-col overflow-hidden px-3 pt-5 pb-2">
+      }
+    >
+      <div
+        className={cn(
+          "flex min-h-0 min-w-0 flex-1 flex-col overflow-hidden px-3 pt-5 pb-2",
+          reveal
+        )}
+      >
         {sources}
       </div>
-      {footer}
+      <div className={cn("contents", reveal)}>{footer}</div>
       <ChatsDialog
         open={chatsOpen}
         onOpenChange={setChatsOpen}
@@ -220,6 +225,6 @@ export const LeftSidebar = memo(function LeftSidebar({
         runStates={runStates}
         unreadThreadIds={unreadThreadIds}
       />
-    </aside>
+    </SidebarFrame>
   )
 })
diff --git a/surfsense_local/frontend/src/features/dashboard/right-panel.tsx b/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
index 636496b8e0..3f2a1e2cf7 100644
--- a/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
@@ -1,7 +1,9 @@
 import type { ReactNode } from "react"
 
 import { ColumnHeader } from "@/components/ui/column-header"
+import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 import { intl } from "@/i18n/intl"
+import { cn } from "@/lib/utils"
 
 // The right column: Studio's generate control sits above the artifact
 // list, and an inspected citation or artifact swaps out for the whole panel.
@@ -19,6 +21,7 @@ export function RightPanel({
   /** Hides the panel; Studio's heading holds it, a preview's header never does. */
   toggle?: ReactNode
 }) {
+  const reveal = useStartupReveal()
   if (inspect) return inspect
 
   return (
@@ -40,9 +43,14 @@ export function RightPanel({
           </h2>
           {toggle}
         </ColumnHeader>
-        <div className="px-3 pt-2">{studio}</div>
+        <div className={cn("px-3 pt-2", reveal)}>{studio}</div>
       </header>
-      <div className="min-h-0 min-w-0 flex-1 overflow-hidden px-3 pb-2">
+      <div
+        className={cn(
+          "min-h-0 min-w-0 flex-1 overflow-hidden px-3 pb-2",
+          reveal
+        )}
+      >
         {artifacts}
       </div>
     </aside>
diff --git a/surfsense_local/frontend/src/features/dashboard/sidebar-frame.tsx b/surfsense_local/frontend/src/features/dashboard/sidebar-frame.tsx
new file mode 100644
index 0000000000..78004f1f25
--- /dev/null
+++ b/surfsense_local/frontend/src/features/dashboard/sidebar-frame.tsx
@@ -0,0 +1,29 @@
+import type { ReactNode } from "react"
+
+import { ColumnHeader } from "@/components/ui/column-header"
+
+/**
+ * The left column's box and its "SurfSense" heading, which never change, so
+ * the startup frame draws them too. `nav` sits under the heading in its header.
+ */
+export function SidebarFrame({
+  nav,
+  children,
+}: {
+  nav?: ReactNode
+  children?: ReactNode
+}) {
+  return (
+    <aside className="flex h-full min-w-0 flex-col border-r bg-background select-none">
+      <header>
+        <ColumnHeader>
+          <h2 className="truncate px-1 brand-wordmark text-xl text-foreground select-none">
+            SurfSense
+          </h2>
+        </ColumnHeader>
+        {nav}
+      </header>
+      {children}
+    </aside>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/startup/dashboard-hint.test.ts b/surfsense_local/frontend/src/features/startup/dashboard-hint.test.ts
new file mode 100644
index 0000000000..cab3a06c78
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/dashboard-hint.test.ts
@@ -0,0 +1,56 @@
+import { afterEach, describe, expect, it } from "vitest"
+
+import {
+  expectsDashboard,
+  forgetDashboard,
+  rememberDashboard,
+} from "./dashboard-hint"
+
+function onDatabase(identity: string | null) {
+  window.surfsense = {
+    databaseIdentity: () => identity,
+  } as unknown as Window["surfsense"]
+}
+
+afterEach(() => {
+  window.surfsense = undefined
+  localStorage.clear()
+})
+
+describe("dashboard hint", () => {
+  it("expects nothing before onboarding has been seen done", () => {
+    onDatabase("a")
+    expect(expectsDashboard()).toBe(false)
+  })
+
+  it("expects the dashboard on the database it was remembered for", () => {
+    onDatabase("a")
+    rememberDashboard()
+    expect(expectsDashboard()).toBe(true)
+  })
+
+  it("expects nothing once the database has been made again", () => {
+    onDatabase("a")
+    rememberDashboard()
+    onDatabase("b")
+    expect(expectsDashboard()).toBe(false)
+  })
+
+  it("expects nothing once onboarding is required again", () => {
+    onDatabase("a")
+    rememberDashboard()
+    forgetDashboard()
+    expect(expectsDashboard()).toBe(false)
+  })
+
+  it("expects nothing while the database does not exist yet", () => {
+    onDatabase(null)
+    rememberDashboard()
+    expect(expectsDashboard()).toBe(false)
+  })
+
+  it("expects nothing without the desktop's database identity", () => {
+    rememberDashboard()
+    expect(expectsDashboard()).toBe(false)
+  })
+})
diff --git a/surfsense_local/frontend/src/features/startup/dashboard-hint.ts b/surfsense_local/frontend/src/features/startup/dashboard-hint.ts
new file mode 100644
index 0000000000..a5bc57bbbd
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/dashboard-hint.ts
@@ -0,0 +1,45 @@
+// Holds the identity of the database whose onboarding was last seen done.
+const DASHBOARD_HINT_KEY = "surfsense:startup-dashboard:v1"
+
+// Docker has no bridge, so it never expects the dashboard: its data volume can
+// be replaced under a browser that still remembers the old one.
+function databaseIdentity(): string | null {
+  try {
+    return window.surfsense?.databaseIdentity?.() ?? null
+  } catch {
+    return null
+  }
+}
+
+/**
+ * Whether startup will open the dashboard, guessed before the API answers so
+ * the loading screen can draw it. Only a hint: wrong, it costs one frame, and
+ * on a database made since it was remembered it never guesses yes.
+ */
+export function expectsDashboard(): boolean {
+  const identity = databaseIdentity()
+  if (identity === null) return false
+  try {
+    return localStorage.getItem(DASHBOARD_HINT_KEY) === identity
+  } catch {
+    return false
+  }
+}
+
+export function rememberDashboard(): void {
+  const identity = databaseIdentity()
+  if (identity === null) return
+  try {
+    localStorage.setItem(DASHBOARD_HINT_KEY, identity)
+  } catch {
+    // Without storage, every launch shows the plain loading screen.
+  }
+}
+
+export function forgetDashboard(): void {
+  try {
+    localStorage.removeItem(DASHBOARD_HINT_KEY)
+  } catch {
+    // Nothing was stored.
+  }
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-frame.tsx b/surfsense_local/frontend/src/features/startup/startup-frame.tsx
new file mode 100644
index 0000000000..fe8956c145
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-frame.tsx
@@ -0,0 +1,77 @@
+import { useState } from "react"
+
+import {
+  THREAD_BODY_CLASS,
+  VIEWPORT_BOX_CLASS,
+  VIEWPORT_CLASS,
+  VIEWPORT_FOOTER_CLASS,
+  VIEWPORT_FOOTER_INNER_CLASS,
+  VIEWPORT_STYLE,
+} from "@/features/chat/chat-viewport-layout"
+import { ComposerOutline } from "@/features/chat/composer-outline"
+import { ConversationColumn } from "@/features/chat/conversation-column"
+import { MessagesSkeleton } from "@/features/chat/messages-skeleton"
+import { ThreadWelcome } from "@/features/chat/thread-welcome"
+import {
+  ChatColumn,
+  DashboardColumns,
+  RightPanelRail,
+  SidebarColumn,
+} from "@/features/dashboard/dashboard-columns"
+import { RightPanel } from "@/features/dashboard/right-panel"
+import { SidebarFrame } from "@/features/dashboard/sidebar-frame"
+
+import { readStartupLayout } from "./startup-layout"
+
+/**
+ * The dashboard's first frame without its data: the same columns, headings
+ * and composer box, drawn by the dashboard's own parts, with nothing to press.
+ * What needs data fades in over it once startup ends.
+ */
+export function StartupFrame() {
+  const [layout] = useState(readStartupLayout)
+  return (
+    <>
+      {/* The workspace rail's column, empty until its workspaces arrive. */}
+      <div />
+      <DashboardColumns
+        sidebarWide={false}
+        rightPanelOpen={layout.rightPanelOpen}
+        rightPanelWide={false}
+      >
+        <SidebarColumn id="startup-sidebar">
+          <SidebarFrame />
+        </SidebarColumn>
+        <ChatColumn>
+          <ConversationColumn>
+            <div className={THREAD_BODY_CLASS}>
+              <div className={VIEWPORT_BOX_CLASS}>
+                <div className={VIEWPORT_CLASS} style={VIEWPORT_STYLE}>
+                  {layout.composer === "center" ? (
+                    <ThreadWelcome
+                      composer={<ComposerOutline placement="center" />}
+                    />
+                  ) : (
+                    <>
+                      <MessagesSkeleton />
+                      <div className={VIEWPORT_FOOTER_CLASS}>
+                        <div className={VIEWPORT_FOOTER_INNER_CLASS}>
+                          <ComposerOutline placement="bottom" />
+                        </div>
+                      </div>
+                    </>
+                  )}
+                </div>
+              </div>
+            </div>
+          </ConversationColumn>
+        </ChatColumn>
+        <RightPanelRail open={layout.rightPanelOpen}>
+          <div className="h-full min-h-0">
+            <RightPanel inspect={null} studio={null} artifacts={null} />
+          </div>
+        </RightPanelRail>
+      </DashboardColumns>
+    </>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-layout.test.ts b/surfsense_local/frontend/src/features/startup/startup-layout.test.ts
new file mode 100644
index 0000000000..70d4baeb9d
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-layout.test.ts
@@ -0,0 +1,32 @@
+import { afterEach, describe, expect, it } from "vitest"
+
+import { readStartupLayout } from "./startup-layout"
+
+afterEach(() => localStorage.clear())
+
+describe("startup layout", () => {
+  it("opens on a new chat with the right panel open when nothing is stored", () => {
+    expect(readStartupLayout()).toEqual({
+      composer: "center",
+      rightPanelOpen: true,
+    })
+  })
+
+  it("puts the composer at the bottom when the last workspace had a thread open", () => {
+    localStorage.setItem("surfsense:last-workspace:v1", "3")
+    localStorage.setItem("surfsense:last-thread:3:v1", "12")
+    expect(readStartupLayout().composer).toBe("bottom")
+  })
+
+  it("centres the composer when the last workspace was on a new chat", () => {
+    localStorage.setItem("surfsense:last-workspace:v1", "3")
+    localStorage.setItem("surfsense:last-thread:3:v1", "new")
+    localStorage.setItem("surfsense:last-thread:4:v1", "12")
+    expect(readStartupLayout().composer).toBe("center")
+  })
+
+  it("keeps the right panel closed when it was closed", () => {
+    localStorage.setItem("surfsense:right-panel:v1", "collapsed")
+    expect(readStartupLayout().rightPanelOpen).toBe(false)
+  })
+})
diff --git a/surfsense_local/frontend/src/features/startup/startup-layout.ts b/surfsense_local/frontend/src/features/startup/startup-layout.ts
new file mode 100644
index 0000000000..5a1837c909
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-layout.ts
@@ -0,0 +1,24 @@
+import { readLastThread } from "@/features/chat/last-thread"
+import { readRightPanelOpen } from "@/features/dashboard/chrome-prefs"
+import { readStoredWorkspaceId } from "@/features/workspaces/last-workspace"
+
+export type StartupLayout = {
+  composer: "center" | "bottom"
+  rightPanelOpen: boolean
+}
+
+/**
+ * How the dashboard's first frame will look, from what it stored when last
+ * open, read with the readers it uses so the loading frame agrees. A stored
+ * source preview is left out: it opens only once the sources have loaded.
+ */
+export function readStartupLayout(): StartupLayout {
+  const workspaceId = readStoredWorkspaceId()
+  return {
+    composer:
+      workspaceId !== null && readLastThread(workspaceId) !== null
+        ? "bottom"
+        : "center",
+    rightPanelOpen: readRightPanelOpen(),
+  }
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-reveal-context.ts b/surfsense_local/frontend/src/features/startup/startup-reveal-context.ts
new file mode 100644
index 0000000000..14f3e81d86
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-reveal-context.ts
@@ -0,0 +1,4 @@
+import { createContext } from "react"
+
+// True while the dashboard's first commit is fading in over the startup frame.
+export const StartupRevealContext = createContext(false)
diff --git a/surfsense_local/frontend/src/features/startup/startup-reveal-provider.tsx b/surfsense_local/frontend/src/features/startup/startup-reveal-provider.tsx
new file mode 100644
index 0000000000..4b594c6500
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-reveal-provider.tsx
@@ -0,0 +1,28 @@
+import { useEffect, useState, type ReactNode } from "react"
+
+import { StartupRevealContext } from "./startup-reveal-context"
+
+/**
+ * Whether the dashboard's first draw fades in over the loading frame. True
+ * through its first commit only, so nothing mounted later fades.
+ */
+export function StartupRevealProvider({
+  reveal,
+  children,
+}: {
+  reveal: boolean
+  children: ReactNode
+}) {
+  const [active, setActive] = useState(reveal)
+  useEffect(() => {
+    if (!active) return
+    // After this commit's own effects, which have read it already.
+    const timer = setTimeout(() => setActive(false), 0)
+    return () => clearTimeout(timer)
+  }, [active])
+  return (
+    <StartupRevealContext.Provider value={active}>
+      {children}
+    </StartupRevealContext.Provider>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx b/surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx
new file mode 100644
index 0000000000..1062484929
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx
@@ -0,0 +1,73 @@
+import { act, cleanup, render, screen } from "@testing-library/react"
+import userEvent from "@testing-library/user-event"
+import { useState } from "react"
+import { afterEach, describe, expect, it, vi } from "vitest"
+
+import { StartupRevealProvider } from "./startup-reveal-provider"
+import { useStartupReveal } from "./use-startup-reveal"
+
+function Region({ name }: { name: string }) {
+  return <div data-testid={name} className={useStartupReveal()} />
+}
+
+function LateRegion() {
+  const [shown, setShown] = useState(false)
+  return shown ? (
+    <Region name="late" />
+  ) : (
+    <button type="button" onClick={() => setShown(true)}>
+      show
+    </button>
+  )
+}
+
+afterEach(() => {
+  cleanup()
+  vi.useRealTimers()
+})
+
+describe("startup reveal", () => {
+  it("fades in what the dashboard first draws after a visible wait", () => {
+    render(
+      <StartupRevealProvider reveal>
+        <Region name="region" />
+      </StartupRevealProvider>
+    )
+    expect(screen.getByTestId("region").className).toContain("ss-reveal")
+  })
+
+  it("draws it at once after a wait too short to see", () => {
+    render(
+      <StartupRevealProvider reveal={false}>
+        <Region name="region" />
+      </StartupRevealProvider>
+    )
+    expect(screen.getByTestId("region").className).not.toContain("ss-reveal")
+  })
+
+  it("stops fading once the fade is over, so later contents arrive as usual", () => {
+    vi.useFakeTimers()
+    render(
+      <StartupRevealProvider reveal>
+        <Region name="region" />
+      </StartupRevealProvider>
+    )
+    act(() => vi.advanceTimersByTime(1000))
+    expect(screen.getByTestId("region").className).not.toContain("ss-reveal")
+  })
+
+  it("never fades what mounts after startup", async () => {
+    render(
+      <StartupRevealProvider reveal>
+        <LateRegion />
+      </StartupRevealProvider>
+    )
+    await userEvent.click(screen.getByRole("button", { name: "show" }))
+    expect(screen.getByTestId("late").className).not.toContain("ss-reveal")
+  })
+
+  it("fades nothing outside startup", () => {
+    render(<Region name="region" />)
+    expect(screen.getByTestId("region").className).not.toContain("ss-reveal")
+  })
+})
diff --git a/surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts b/surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts
new file mode 100644
index 0000000000..39ec02afbb
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts
@@ -0,0 +1,77 @@
+import { describe, expect, it } from "vitest"
+
+// The startup screen ships in the first chunk, before the dashboard's. If it
+// reached the chat runtime or a feature's API client, that code would load
+// before anything is drawn and the frame would arrive late.
+const FORBIDDEN = [
+  /^@assistant-ui\//,
+  /^@tanstack\/react-query/,
+  /^\/src\/features\/.+\/api\.ts$/,
+]
+
+// Every source file's text, keyed from the project root: "/src/…".
+const SOURCES = import.meta.glob<string>(
+  ["/src/**/*.{ts,tsx}", "!/src/**/*.test.{ts,tsx}"],
+  { query: "?raw", import: "default", eager: true }
+)
+
+function resolveLocal(from: string, specifier: string): string | null {
+  const base = specifier.startsWith("@/")
+    ? `/src/${specifier.slice(2)}`
+    : new URL(specifier, `file://${from}`).pathname
+  for (const candidate of [
+    base,
+    `${base}.ts`,
+    `${base}.tsx`,
+    `${base}/index.ts`,
+    `${base}/index.tsx`,
+  ]) {
+    if (candidate in SOURCES) return candidate
+  }
+  return null
+}
+
+// Every module the file loads at run time; type-only imports are erased.
+function reachable(entry: string): string[] {
+  const seen = new Set<string>()
+  const queue = [entry]
+  const reached: string[] = []
+  while (queue.length > 0) {
+    const file = queue.pop()!
+    if (seen.has(file)) continue
+    seen.add(file)
+    const imports = (SOURCES[file] ?? "").matchAll(
+      /^\s*import\s+(?!type\b)(?:[^"';]*?\sfrom\s+)?["']([^"']+)["']/gm
+    )
+    for (const [, specifier] of imports) {
+      if (specifier.startsWith(".") || specifier.startsWith("@/")) {
+        const local = resolveLocal(file, specifier)
+        if (local) {
+          reached.push(local)
+          queue.push(local)
+        }
+      } else {
+        reached.push(specifier)
+      }
+    }
+  }
+  return reached
+}
+
+describe("startup screen's imports", () => {
+  it("are walked deep enough to find the runtime where it is used", () => {
+    const reached = reachable("/src/features/chat/thread-panel.tsx")
+    expect(reached.some((module) => FORBIDDEN[0]!.test(module))).toBe(true)
+    expect(reached.some((module) => FORBIDDEN[2]!.test(module))).toBe(true)
+  })
+
+  it("reach neither the chat runtime nor any feature's API client", () => {
+    const reached = reachable("/src/features/startup/startup-screen.tsx")
+    expect(reached.length).toBeGreaterThan(10)
+    expect(
+      reached.filter((module) =>
+        FORBIDDEN.some((pattern) => pattern.test(module))
+      )
+    ).toEqual([])
+  })
+})
diff --git a/surfsense_local/frontend/src/features/startup/startup-screen.tsx b/surfsense_local/frontend/src/features/startup/startup-screen.tsx
new file mode 100644
index 0000000000..2a596cf719
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-screen.tsx
@@ -0,0 +1,25 @@
+import { DashboardShell } from "@/features/dashboard/dashboard-shell"
+import { intl } from "@/i18n/intl"
+
+import { StartupFrame } from "./startup-frame"
+
+/**
+ * What shows while startup waits on the API: the dashboard's frame when the
+ * dashboard is expected next, else the window's plain background, so
+ * onboarding never opens over a glimpse of the dashboard.
+ */
+export function StartupScreen({ frame }: { frame: boolean }) {
+  return (
+    <DashboardShell
+      role="status"
+      aria-busy="true"
+      aria-label={intl.formatMessage({
+        id: "app_bootstrap_loader_aria",
+        defaultMessage: "Starting SurfSense",
+      })}
+      className="select-none"
+    >
+      {frame ? <StartupFrame /> : null}
+    </DashboardShell>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/startup/use-startup-reveal.ts b/surfsense_local/frontend/src/features/startup/use-startup-reveal.ts
new file mode 100644
index 0000000000..f275b20dd7
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/use-startup-reveal.ts
@@ -0,0 +1,21 @@
+import { useContext, useEffect, useState } from "react"
+
+import { StartupRevealContext } from "./startup-reveal-context"
+
+/** `.ss-reveal` in `index.css` runs 200 ms; the class is dropped after it ends. */
+const REVEAL_CLEAR_MS = 400
+
+/**
+ * The class for a container whose contents fade in at startup while the
+ * container stays put: the loading frame drew the container already.
+ */
+export function useStartupReveal(): string {
+  const reveal = useContext(StartupRevealContext)
+  const [fading, setFading] = useState(reveal)
+  useEffect(() => {
+    if (!fading) return
+    const timer = setTimeout(() => setFading(false), REVEAL_CLEAR_MS)
+    return () => clearTimeout(timer)
+  }, [fading])
+  return fading ? "ss-reveal" : ""
+}
diff --git a/surfsense_local/frontend/src/features/updates/stub-bridge.ts b/surfsense_local/frontend/src/features/updates/stub-bridge.ts
index f6b701b65e..dc6c504888 100644
--- a/surfsense_local/frontend/src/features/updates/stub-bridge.ts
+++ b/surfsense_local/frontend/src/features/updates/stub-bridge.ts
@@ -18,6 +18,7 @@ export function stubUpdateBridge(initial: {
   window.surfsense = {
     apiUrl: "",
     platform: "darwin",
+    databaseIdentity: () => null,
     openDocument: vi.fn(),
     revealDocument: vi.fn(),
     updates: {
diff --git a/surfsense_local/frontend/src/features/workspaces/last-workspace.ts b/surfsense_local/frontend/src/features/workspaces/last-workspace.ts
new file mode 100644
index 0000000000..93439aa32b
--- /dev/null
+++ b/surfsense_local/frontend/src/features/workspaces/last-workspace.ts
@@ -0,0 +1,19 @@
+const LAST_WORKSPACE_KEY = "surfsense:last-workspace:v1"
+
+/** The workspace chosen last, before the list says whether it still exists. */
+export function readStoredWorkspaceId(): number | null {
+  try {
+    const stored = Number(localStorage.getItem(LAST_WORKSPACE_KEY))
+    return Number.isInteger(stored) && stored > 0 ? stored : null
+  } catch {
+    return null
+  }
+}
+
+export function rememberWorkspace(id: number) {
+  try {
+    localStorage.setItem(LAST_WORKSPACE_KEY, String(id))
+  } catch {
+    // Selection still works for this session when storage is unavailable.
+  }
+}
diff --git a/surfsense_local/frontend/src/features/workspaces/use-workspaces.ts b/surfsense_local/frontend/src/features/workspaces/use-workspaces.ts
index 32c84b12d1..0af17fabf1 100644
--- a/surfsense_local/frontend/src/features/workspaces/use-workspaces.ts
+++ b/surfsense_local/frontend/src/features/workspaces/use-workspaces.ts
@@ -9,30 +9,17 @@ import {
   renameWorkspace,
   type Workspace,
 } from "./api"
-
-const LAST_WORKSPACE_KEY = "surfsense:last-workspace:v1"
+import { readStoredWorkspaceId, rememberWorkspace } from "./last-workspace"
 
 function readLastWorkspaceId(workspaces: Workspace[]): number | null {
-  try {
-    const stored = Number(localStorage.getItem(LAST_WORKSPACE_KEY))
-    if (workspaces.some((workspace) => workspace.id === stored)) {
-      return stored
-    }
-  } catch {
-    // Storage can be disabled; the first workspace is a safe fallback.
+  const stored = readStoredWorkspaceId()
+  if (workspaces.some((workspace) => workspace.id === stored)) {
+    return stored
   }
   // Null when the list is empty, so nothing is active.
   return workspaces[0]?.id ?? null
 }
 
-function rememberWorkspace(id: number) {
-  try {
-    localStorage.setItem(LAST_WORKSPACE_KEY, String(id))
-  } catch {
-    // Selection still works for this session when storage is unavailable.
-  }
-}
-
 function messageFrom(error: unknown) {
   return error instanceof Error
     ? error.message
diff --git a/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx b/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx
index 158870830a..2f357d2fd0 100644
--- a/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx
+++ b/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx
@@ -40,6 +40,7 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip"
+import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 import { intl } from "@/i18n/intl"
 import { cn } from "@/lib/utils"
 
@@ -172,6 +173,7 @@ export function WorkspaceRail({
   const [namingOpen, setNamingOpen] = useState(false)
   const [deleting, setDeleting] = useState<Workspace | null>(null)
   const [deleteOpen, setDeleteOpen] = useState(false)
+  const reveal = useStartupReveal()
 
   const openNaming = (target: Workspace | "new") => {
     setNaming(target)
@@ -180,7 +182,10 @@ export function WorkspaceRail({
 
   return (
     <nav
-      className="flex h-full flex-col items-center bg-app-shell py-3 text-sidebar-foreground"
+      className={cn(
+        "flex h-full flex-col items-center bg-app-shell py-3 text-sidebar-foreground",
+        reveal
+      )}
       aria-label={intl.formatMessage({
         id: "workspaces_rail_aria",
         defaultMessage: "Workspaces",
diff --git a/surfsense_local/frontend/src/index.css b/surfsense_local/frontend/src/index.css
index db889520d8..d58f99b9b4 100644
--- a/surfsense_local/frontend/src/index.css
+++ b/surfsense_local/frontend/src/index.css
@@ -413,53 +413,22 @@
   );
 }
 
-/* The startup loader: the logo fills with a rising, rippling tide
- * (`app/logo-fill-loader.tsx`). Transforms and opacity only, so the
- * compositor runs it however busy the first load keeps the page. It fills,
- * holds, fades back to faint while full, and rises again. */
-@keyframes ss-logo-fill-rise {
-  0% {
-    transform: translateY(100%);
-    opacity: 1;
-  }
-  70%,
-  85% {
-    transform: translateY(-0.5rem);
-    opacity: 1;
-  }
-  100% {
-    transform: translateY(-0.5rem);
+/* Startup: the dashboard's contents fade in where the loading frame already
+ * drew their containers (`features/startup/startup-reveal.tsx`). The
+ * children fade, never the container, so a drawn box such as the composer's
+ * stays put. Opacity only: nothing moves out of the frame's place. */
+@keyframes ss-reveal-in {
+  from {
     opacity: 0;
   }
 }
 
-@keyframes ss-logo-fill-wave {
-  to {
-    transform: translateX(-50%);
-  }
-}
-
-/* `--ss-logo-fill-offset` is a negative delay set from the page clock, so a
- * loader that mounts mid-startup joins the tide where it is instead of
- * starting it again from empty. The wave's cycle divides the rise's, so one
- * offset keeps both in step. Keep 2.2s in step with `FILL_CYCLE_MS`. */
-.ss-logo-fill-liquid {
-  animation: ss-logo-fill-rise 2.2s cubic-bezier(0.45, 0, 0.2, 1) infinite;
-  animation-delay: var(--ss-logo-fill-offset, 0ms);
-}
-
-.ss-logo-fill-wave {
-  animation: ss-logo-fill-wave 1.1s linear infinite;
-  animation-delay: var(--ss-logo-fill-offset, 0ms);
+.ss-reveal > * {
+  animation: ss-reveal-in 200ms cubic-bezier(0.215, 0.61, 0.355, 1);
 }
 
 @media (prefers-reduced-motion: reduce) {
-  .ss-logo-fill-liquid {
-    animation: none;
-    transform: translateY(-0.5rem);
-  }
-
-  .ss-logo-fill-wave {
+  .ss-reveal > * {
     animation: none;
   }
 }
diff --git a/surfsense_local/frontend/src/lib/api.ts b/surfsense_local/frontend/src/lib/api.ts
index 6ce238cc69..54e5c18510 100644
--- a/surfsense_local/frontend/src/lib/api.ts
+++ b/surfsense_local/frontend/src/lib/api.ts
@@ -9,6 +9,8 @@ declare global {
     surfsense?: {
       apiUrl: string
       platform: string
+      // Changes when the database is made again; null before it exists.
+      databaseIdentity: () => string | null
       openDocument: (workspaceId: number, documentId: number) => Promise<string>
       revealDocument: (
         workspaceId: number,

From 80a863e9bd35960179475537d9f725027ff502b8 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 01:09:26 +0530
Subject: [PATCH 02/12] feat(local): take typing in the startup frame, with @
 mentions

A new chat's box takes typing while the app loads and hands the text,
caret and focus to the composer, through a store AppBootstrap provides.
One @ rule serves the composer's plugin list and the frame, which shows
a skeleton of the list when new chats open in Agentic. An open thread's
frame shows its messages' skeleton and gets its composer with them.
---
 docs/architecture/overview.md                 |   2 +-
 surfsense_local/frontend/package.json         |   1 +
 surfsense_local/frontend/pnpm-lock.yaml       |   3 +
 .../src/app/app-bootstrap-startup.test.tsx    | 249 ++++++++++++++++--
 .../frontend/src/app/app-bootstrap.tsx        |  63 +++--
 .../src/features/chat/chat-composer.test.tsx  | 152 +++++++++--
 .../src/features/chat/chat-composer.tsx       | 123 +++++----
 .../src/features/chat/chat-viewport-layout.ts |  19 --
 .../src/features/chat/chat-viewport.tsx       |  52 ++--
 .../src/features/chat/composer-layout.ts      |  33 ---
 .../src/features/chat/composer-outline.tsx    |  36 ---
 .../src/features/chat/composer-surface.tsx    | 101 +++++++
 .../src/features/chat/composer-wording.ts     |  17 ++
 .../src/features/chat/conversation-column.tsx |  13 +-
 .../src/features/chat/conversation-scroll.tsx |  55 ++++
 .../chat/modes/next-new-chat-mode-store.ts    |  22 ++
 .../chat/modes/next-new-chat-mode.test.ts     |  57 ++++
 .../features/chat/modes/next-new-chat-mode.ts |  13 +
 .../src/features/chat/thread-panel.tsx        |   8 +-
 .../src/features/dashboard/dashboard-page.tsx |   9 +-
 .../src/features/dashboard/left-sidebar.tsx   |  12 +-
 .../src/features/dashboard/right-panel.tsx    |   7 +-
 .../plugins/in-thread/mention-list.tsx        |  68 +++++
 .../plugins/in-thread/mention-matcher.test.ts |  26 ++
 .../plugins/in-thread/mention-matcher.ts      |  30 +++
 .../plugins/in-thread/mention-queries.ts      |  21 ++
 .../plugins/in-thread/plugin-mentions.tsx     |  63 ++---
 .../src/features/startup/startup-composer.tsx | 108 ++++++++
 .../startup/startup-draft-handover.tsx        |  53 ++++
 .../src/features/startup/startup-frame.tsx    |  51 ++--
 .../features/startup/startup-input-context.ts |  10 +
 .../features/startup/startup-input.test.ts    |  44 ++++
 .../src/features/startup/startup-input.ts     |  43 +++
 .../features/startup/startup-layout.test.ts   |   1 +
 .../src/features/startup/startup-layout.ts    |   4 +
 .../features/startup/startup-reveal.test.tsx  |  21 +-
 .../startup/startup-screen-imports.test.ts    |  54 +++-
 .../src/features/startup/startup-screen.tsx   |  18 +-
 .../src/features/startup/use-startup-draft.ts |  16 ++
 .../features/startup/use-startup-reveal.ts    |  34 ++-
 .../features/workspaces/workspace-rail.tsx    |   3 +-
 surfsense_local/frontend/src/index.css        |   5 +-
 42 files changed, 1381 insertions(+), 339 deletions(-)
 delete mode 100644 surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts
 delete mode 100644 surfsense_local/frontend/src/features/chat/composer-layout.ts
 delete mode 100644 surfsense_local/frontend/src/features/chat/composer-outline.tsx
 create mode 100644 surfsense_local/frontend/src/features/chat/composer-surface.tsx
 create mode 100644 surfsense_local/frontend/src/features/chat/composer-wording.ts
 create mode 100644 surfsense_local/frontend/src/features/chat/conversation-scroll.tsx
 create mode 100644 surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode-store.ts
 create mode 100644 surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.test.ts
 create mode 100644 surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts
 create mode 100644 surfsense_local/frontend/src/features/plugins/in-thread/mention-list.tsx
 create mode 100644 surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.test.ts
 create mode 100644 surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.ts
 create mode 100644 surfsense_local/frontend/src/features/plugins/in-thread/mention-queries.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-composer.tsx
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-draft-handover.tsx
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-input-context.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-input.test.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/startup-input.ts
 create mode 100644 surfsense_local/frontend/src/features/startup/use-startup-draft.ts

diff --git a/docs/architecture/overview.md b/docs/architecture/overview.md
index dd984af390..da2c9ab41a 100644
--- a/docs/architecture/overview.md
+++ b/docs/architecture/overview.md
@@ -110,7 +110,7 @@ The rules that keep the processes out of each other's way:
 - API calls go through `request()` in [`lib/api.ts`](../../surfsense_local/frontend/src/lib/api.ts), except the Studio viewers, which fetch an artifact's file bytes directly; an `<img>`, `<audio>` or download link takes its absolute address from `apiUrl()`. `request()` prefixes the address, turns an error body into an `ApiError` with its `code`, and when a user action is refused with `403 egress_disabled` it asks for consent and retries once.
 - TanStack Query holds most server state: threads and messages, the citation panel, the model catalog and settings, the Studio formats and artifact list, and an open artifact with what its viewer reads. An `artifacts` event invalidates the list and every open artifact under one key prefix ([`features/studio/query-keys.ts`](../../surfsense_local/frontend/src/features/studio/query-keys.ts)), so a regenerated artifact that is open shows its new generation. The sources list is component state, reloaded on the events above.
 - There is no router. [`app/app-bootstrap.tsx`](../../surfsense_local/frontend/src/app/app-bootstrap.tsx) asks `GET /llm/onboarding`, then renders either onboarding ([`local-models/selection.md`](local-models/selection.md)) or the lazily loaded dashboard: the workspace rail, threads and sources on the left, the conversation in the middle, and Studio, artifacts and the citation panel on the right. The left and right columns resize by dragging their inner edge, and the conversation keeps at least 520 px ([studio](studio.md#the-studio-panel)).
-- While startup waits on the API, [`features/startup/`](../../surfsense_local/frontend/src/features/startup/) draws the dashboard's constant parts, with nothing to press: the columns at their saved widths, the "SurfSense" and "Studio" headings, and the composer's empty box where the last thread put it. The dashboard draws the same parts with the same components, so nothing moves when it takes over; its contents fade in over 200 ms, or appear at once if the wait was under 150 ms or motion is reduced. The frame is drawn only when onboarding was last seen done on the same database, which Electron names by the database file's identity (`window.surfsense.databaseIdentity()`), so a database made since, a first launch and onboarding get the plain background. Docker has no bridge and always gets the plain background. A test keeps the frame's imports off the chat runtime and every feature's `api.ts`, so it stays in the first chunk.
+- While startup waits on the API, [`features/startup/`](../../surfsense_local/frontend/src/features/startup/) draws the dashboard's constant parts, with nothing to press: the columns at their saved widths, the "SurfSense" and "Studio" headings, and either a new chat's box or an open thread's message skeleton, as the last view was; a thread's composer arrives with its messages. The new chat's box takes typing but sends nothing, and the text goes to the dashboard's composer through a store `AppBootstrap` owns and provides to both (`startup-input`). The composer sets the text with `composer.setText`, then, once it shows, places the caret and tells the `@` trigger where it is through assistant-ui's trigger context (`resource.setCursorPosition`): the list tracks its own cursor, apart from the box's. Startup waits for an input method's composition to end first. One `@` rule (`mention-matcher.ts`), given to assistant-ui as its matcher, decides in both places whether the caret is in a mention. When new chats open in Agentic, which the dashboard stores for the next launch, `@` in the box shows a skeleton of the plugin list, which startup starts loading early. The dashboard draws the same parts with the same components (`composer-surface.tsx`, `conversation-scroll.tsx`, `mention-list.tsx`, rendered around assistant-ui's parts with Base UI's `useRender`), and both text boxes grow with the same autosizing textarea (`react-textarea-autosize`, which assistant-ui's input is built on), so nothing moves when it takes over; its contents fade in over 200 ms, or appear at once if the wait was under 150 ms or motion is reduced. The frame is drawn only when onboarding was last seen done on the same database, which Electron names by the database file's identity (`window.surfsense.databaseIdentity()`), so a database made since, a first launch and onboarding get the plain background. Docker has no bridge and always gets the plain background. A test walks the frame's imports with TypeScript's parser and keeps them off the chat runtime and every feature's `api.ts`, so it stays in the first chunk. No test measures the two frames' layout in a browser yet.
 - Code is grouped by feature under `src/features/`. Each HTTP contract lives in one feature's `api.ts`, next to its hooks and components, and other features import it from there.
 
 ## Codemap
diff --git a/surfsense_local/frontend/package.json b/surfsense_local/frontend/package.json
index 8b37c9cd85..f86fe63bb8 100644
--- a/surfsense_local/frontend/package.json
+++ b/surfsense_local/frontend/package.json
@@ -47,6 +47,7 @@
     "react-data-grid": "7.0.0-beta.61",
     "react-dom": "^19.2.6",
     "react-intl": "12.1.3",
+    "react-textarea-autosize": "^8.5.9",
     "shadcn": "^4.20.1",
     "shiki": "3.23.0",
     "sonner": "^2.0.8",
diff --git a/surfsense_local/frontend/pnpm-lock.yaml b/surfsense_local/frontend/pnpm-lock.yaml
index e08f4aa4fc..1a75d267c6 100644
--- a/surfsense_local/frontend/pnpm-lock.yaml
+++ b/surfsense_local/frontend/pnpm-lock.yaml
@@ -92,6 +92,9 @@ importers:
       react-intl:
         specifier: 12.1.3
         version: 12.1.3(@types/react@19.2.18)(react@19.2.8)
+      react-textarea-autosize:
+        specifier: ^8.5.9
+        version: 8.5.9(@types/react@19.2.18)(react@19.2.8)
       shadcn:
         specifier: ^4.20.1
         version: 4.20.1(typescript@6.0.3)
diff --git a/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx b/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
index 8009211983..f7f148c2f6 100644
--- a/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
+++ b/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
@@ -1,11 +1,20 @@
-import { cleanup, screen, within } from "@testing-library/react"
+import {
+  cleanup,
+  fireEvent,
+  screen,
+  waitFor,
+  within,
+} from "@testing-library/react"
+import userEvent from "@testing-library/user-event"
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest"
 
 import { TooltipProvider } from "@/components/ui/tooltip"
+import { writeNextNewChatMode } from "@/features/chat/modes/next-new-chat-mode-store"
 import {
   expectsDashboard,
   rememberDashboard,
 } from "@/features/startup/dashboard-hint"
+import { readStartupLayout } from "@/features/startup/startup-layout"
 import { render } from "@/test-utils"
 import { AppBootstrap } from "./app-bootstrap"
 
@@ -27,15 +36,41 @@ function pendingStartup() {
 
 // A completed install with one workspace and no model; `onboardingAfterMs`
 // holds the first answer back, as a slow start does.
-function completedInstall({ onboardingAfterMs = 0 } = {}) {
+function completedInstall({
+  onboardingAfterMs = 0,
+  selection = null as Record<string, unknown> | null,
+  // Holds the first answer until the test resolves it.
+  gate = null as Promise<void> | null,
+} = {}) {
+  const requested: string[] = []
   vi.stubGlobal(
     "fetch",
     vi.fn(async (input: RequestInfo | URL) => {
       const path = String(input)
+      requested.push(path)
       if (path === "/llm/onboarding") {
         await new Promise((resolve) => setTimeout(resolve, onboardingAfterMs))
+        await gate
         return Response.json({ completed: true })
       }
+      if (path === "/plugins/third-party") {
+        return Response.json({ allowed: true })
+      }
+      if (path === "/llm/selection/text_gen" && selection) {
+        return Response.json(selection)
+      }
+      if (path === "/llm/connections/1/models" && selection) {
+        return Response.json([
+          {
+            connection_id: 1,
+            connection_label: "Remote",
+            name: selection.name,
+            types: ["text_gen"],
+            capability_source: "catalog",
+            selectable_for: ["text_gen"],
+          },
+        ])
+      }
       if (path === "/workspaces") {
         return Response.json([
           {
@@ -48,6 +83,7 @@ function completedInstall({ onboardingAfterMs = 0 } = {}) {
       }
       if (
         path === "/llm/providers" ||
+        path === "/plugins" ||
         path === "/workspaces/1/chat/threads" ||
         path === "/workspaces/1/documents?document_type=FILE&document_type=NOTE"
       ) {
@@ -56,6 +92,34 @@ function completedInstall({ onboardingAfterMs = 0 } = {}) {
       return Response.json({ detail: "not found" }, { status: 404 })
     })
   )
+  return requested
+}
+
+function gated() {
+  let release = () => {}
+  const gate = new Promise<void>((resolve) => {
+    release = resolve
+  })
+  return { gate, release }
+}
+
+// The loading screen, labelled for assistive tech while it is busy.
+function startupScreen() {
+  return screen.getByRole("main", STARTING)
+}
+
+function frameBox() {
+  return within(startupScreen()).getByRole<HTMLTextAreaElement>("textbox", {
+    name: "Message",
+  })
+}
+
+async function dashboardRail() {
+  return screen.findByRole(
+    "navigation",
+    { name: "Workspaces" },
+    { timeout: 10_000 }
+  )
 }
 
 function renderApp() {
@@ -94,14 +158,21 @@ afterEach(() => {
 })
 
 describe("startup screen", () => {
+  it("says the app is starting, for assistive tech", () => {
+    pendingStartup()
+    renderApp()
+
+    expect(startupScreen().getAttribute("aria-busy")).toBe("true")
+    expect(screen.getByRole("status").textContent).toBe("Starting SurfSense")
+  })
+
   it("is blank until onboarding has been seen done on this database", () => {
     onDatabase("a")
     pendingStartup()
     renderApp()
 
-    const screenBox = screen.getByRole("status", STARTING)
-    expect(screenBox.textContent).toBe("")
-    expect(within(screenBox).queryByRole("heading")).toBeNull()
+    expect(within(startupScreen()).queryByRole("heading")).toBeNull()
+    expect(within(startupScreen()).queryByRole("textbox")).toBeNull()
   })
 
   it("is blank without the desktop, whose data can change under the browser", () => {
@@ -109,7 +180,7 @@ describe("startup screen", () => {
     pendingStartup()
     renderApp()
 
-    expect(screen.getByRole("status", STARTING).textContent).toBe("")
+    expect(within(startupScreen()).queryByRole("heading")).toBeNull()
   })
 
   it("draws the dashboard's constant parts, and nothing to press, once onboarding is done", () => {
@@ -118,39 +189,150 @@ describe("startup screen", () => {
     pendingStartup()
     renderApp()
 
-    const frame = screen.getByRole("status", STARTING)
-    expect(frame.getAttribute("aria-busy")).toBe("true")
+    const frame = startupScreen()
     expect(
       within(frame).getByRole("heading", { name: "SurfSense" })
     ).toBeTruthy()
     expect(within(frame).getByRole("heading", { name: "Studio" })).toBeTruthy()
     expect(within(frame).queryAllByRole("button")).toEqual([])
-    expect(within(frame).queryByRole("textbox")).toBeNull()
   })
 
-  it("puts the composer where the dashboard will: centred on a new chat", () => {
+  it("holds an open thread's place with its messages' skeleton, and no composer until it loads", () => {
     onDatabase("a")
     rememberDashboard()
+    localStorage.setItem("surfsense:last-workspace:v1", "1")
+    localStorage.setItem("surfsense:last-thread:1:v1", "10")
     pendingStartup()
     renderApp()
 
-    const frame = screen.getByRole("status", STARTING)
-    expect(frame.querySelector("[data-composer-outline=center]")).toBeTruthy()
-    expect(frame.querySelector("[data-composer-outline=bottom]")).toBeNull()
+    const frame = startupScreen()
+    expect(frame.querySelector("[data-slot=skeleton]")).toBeTruthy()
+    expect(within(frame).queryByRole("textbox")).toBeNull()
   })
+})
 
-  it("puts the composer where the dashboard will: at the bottom of an open thread", () => {
+describe("typing while the app loads", () => {
+  function onNewChat() {
     onDatabase("a")
     rememberDashboard()
-    localStorage.setItem("surfsense:last-workspace:v1", "1")
-    localStorage.setItem("surfsense:last-thread:1:v1", "10")
+  }
+
+  it("takes typing in a new chat's box, focused from the start", async () => {
+    onNewChat()
+    pendingStartup()
+    renderApp()
+
+    expect(document.activeElement).toBe(frameBox())
+    await userEvent.type(frameBox(), "find jobs")
+    expect(frameBox().value).toBe("find jobs")
+  })
+
+  it("sends nothing and adds no line on Enter, with nowhere to send yet", async () => {
+    onNewChat()
     pendingStartup()
     renderApp()
 
-    const frame = screen.getByRole("status", STARTING)
-    expect(frame.querySelector("[data-composer-outline=bottom]")).toBeTruthy()
-    expect(frame.querySelector("[data-composer-outline=center]")).toBeNull()
+    await userEvent.type(frameBox(), "find jobs{Enter}")
+    expect(frameBox().value).toBe("find jobs")
   })
+
+  it("carries what was typed into the dashboard's composer", async () => {
+    onNewChat()
+    const { gate, release } = gated()
+    completedInstall({ gate })
+    renderApp()
+
+    await userEvent.type(frameBox(), "find jobs")
+    release()
+    await dashboardRail()
+    await waitFor(() =>
+      expect(
+        screen.getByRole<HTMLTextAreaElement>("textbox", { name: "Message" })
+          .value
+      ).toBe("find jobs")
+    )
+  }, 15_000)
+
+  it("shows the plugin list loading on `@` when new chats open in Agentic", async () => {
+    onNewChat()
+    writeNextNewChatMode("agentic")
+    pendingStartup()
+    renderApp()
+
+    await userEvent.type(frameBox(), "find jobs with @ind")
+    // A busy group: nothing to pick until the plugins arrive.
+    const list = within(startupScreen()).getByRole("group", {
+      name: "Plugins",
+    })
+    expect(list.getAttribute("aria-busy")).toBe("true")
+    expect(list.querySelector("[data-slot=skeleton]")).toBeTruthy()
+  })
+
+  it("shows no plugin list on `@` when new chats open in Q&A", async () => {
+    onNewChat()
+    writeNextNewChatMode("basic")
+    pendingStartup()
+    renderApp()
+
+    await userEvent.type(frameBox(), "find jobs with @ind")
+    expect(
+      within(startupScreen()).queryByRole("group", { name: "Plugins" })
+    ).toBeNull()
+  })
+
+  it("closes the plugin list once the caret leaves the mention", async () => {
+    onNewChat()
+    writeNextNewChatMode("agentic")
+    pendingStartup()
+    renderApp()
+
+    await userEvent.type(frameBox(), "ask @goo about it")
+    expect(
+      within(startupScreen()).queryByRole("group", { name: "Plugins" })
+    ).toBeNull()
+  })
+
+  it("waits for a composition in progress before the dashboard takes the box", async () => {
+    onNewChat()
+    const { gate, release } = gated()
+    completedInstall({ gate })
+    renderApp()
+
+    fireEvent.compositionStart(frameBox())
+    release()
+    await new Promise((resolve) => setTimeout(resolve, 300))
+    expect(startupScreen().getAttribute("aria-busy")).toBe("true")
+    fireEvent.compositionEnd(frameBox())
+    await dashboardRail()
+  }, 15_000)
+
+  it("starts loading the plugins before the dashboard asks for them", async () => {
+    onNewChat()
+    const requested = completedInstall()
+    renderApp()
+
+    await dashboardRail()
+    await waitFor(() =>
+      expect(requested).toContain("/workspaces/1/chat/threads")
+    )
+    expect(requested.indexOf("/plugins")).toBeGreaterThan(-1)
+    expect(requested.indexOf("/plugins")).toBeLessThan(
+      requested.indexOf("/workspaces/1/chat/threads")
+    )
+  }, 15_000)
+
+  it("keeps the dashboard composer's text solid while its controls fade in", async () => {
+    onNewChat()
+    completedInstall({ onboardingAfterMs: 250 })
+    renderApp()
+
+    await dashboardRail()
+    expect(
+      screen
+        .getByRole("textbox", { name: "Message" })
+        .hasAttribute("data-reveal-steady")
+    ).toBe(true)
+  }, 15_000)
 })
 
 describe("startup's memory of onboarding", () => {
@@ -167,6 +349,35 @@ describe("startup's memory of onboarding", () => {
     expect(expectsDashboard()).toBe(true)
   }, 15_000)
 
+  it("keeps the mode the next new chat opens in, for the frame's mentions", async () => {
+    onDatabase("a")
+    completedInstall({
+      selection: {
+        model_type: "text_gen",
+        provider: "openai_compatible",
+        connection_id: 1,
+        name: "anthropic/claude-fable-5",
+        updated_at: "2026-11-01T00:00:00Z",
+        capability: {
+          modes: {
+            agentic_allowed: true,
+            blocked: null,
+            default_mode: "agentic",
+            reason: { code: "assumed", values: {} },
+          },
+        },
+      },
+    })
+    renderApp()
+
+    await screen.findByRole(
+      "navigation",
+      { name: "Workspaces" },
+      { timeout: 10_000 }
+    )
+    expect(readStartupLayout().mentions).toBe(true)
+  }, 15_000)
+
   it("is dropped once the API says onboarding is needed again", async () => {
     onDatabase("a")
     rememberDashboard()
diff --git a/surfsense_local/frontend/src/app/app-bootstrap.tsx b/surfsense_local/frontend/src/app/app-bootstrap.tsx
index e5ff1fa6a8..a6cceb6911 100644
--- a/surfsense_local/frontend/src/app/app-bootstrap.tsx
+++ b/surfsense_local/frontend/src/app/app-bootstrap.tsx
@@ -1,3 +1,4 @@
+import { useQueryClient, type QueryClient } from "@tanstack/react-query"
 import { useEffect, useState } from "react"
 
 import { Alert, AlertDescription, AlertTitle } from "@/components/ui/alert"
@@ -13,6 +14,10 @@ import {
 } from "@/features/models/selection/availability"
 import { getOnboardingStatus } from "@/features/onboarding/api"
 import { OnboardingPage } from "@/features/onboarding/onboarding-page"
+import {
+  pluginListQuery,
+  thirdPartyQuery,
+} from "@/features/plugins/in-thread/mention-queries"
 import { intl } from "@/i18n/intl"
 
 import { listWorkspaces, type Workspace } from "@/features/workspaces/api"
@@ -22,6 +27,11 @@ import {
   forgetDashboard,
   rememberDashboard,
 } from "@/features/startup/dashboard-hint"
+import {
+  createStartupInput,
+  type StartupInput,
+} from "@/features/startup/startup-input"
+import { StartupInputContext } from "@/features/startup/startup-input-context"
 import { StartupRevealProvider } from "@/features/startup/startup-reveal-provider"
 import { StartupScreen } from "@/features/startup/startup-screen"
 
@@ -80,7 +90,7 @@ function messageFrom(error: unknown) {
       })
 }
 
-async function fetchBootstrapState(): Promise<Fetched> {
+async function fetchBootstrapState(queryClient: QueryClient): Promise<Fetched> {
   // Fetched alongside the data rather than after it: one wait, one loader.
   const dashboard = loadDashboard()
   // Handled where it is awaited; this only keeps an early exit from leaving
@@ -93,6 +103,10 @@ async function fetchBootstrapState(): Promise<Fetched> {
       return { status: "onboarding-required" }
     }
     rememberDashboard()
+    // For a mention typed during startup: the list is ready with the composer.
+    // Never awaited, and a failure leaves the list to load as it would.
+    void queryClient.prefetchQuery(pluginListQuery)
+    void queryClient.prefetchQuery(thirdPartyQuery)
     const [selection, workspaces] = await Promise.all([
       getGenerationSelection(),
       listWorkspaces(),
@@ -115,17 +129,33 @@ async function fetchBootstrapState(): Promise<Fetched> {
 }
 
 export function AppBootstrap() {
+  // What is typed into the frame's box, for the composer that replaces it.
+  const [startupInput] = useState(createStartupInput)
+  return (
+    <StartupInputContext.Provider value={startupInput}>
+      <BootstrapScreens startupInput={startupInput} />
+    </StartupInputContext.Provider>
+  )
+}
+
+function BootstrapScreens({ startupInput }: { startupInput: StartupInput }) {
+  // The frame's box may be mid-composition, which a handover would cut.
+  const afterComposing = (next: Fetched) =>
+    startupInput.composingEnded().then(() => next)
+  const queryClient = useQueryClient()
   const [state, setState] = useState<BootstrapState>(startLoading)
 
   useEffect(() => {
     let active = true
-    void fetchBootstrapState().then((next) => {
-      if (active) setState((current) => settle(current, next))
-    })
+    void fetchBootstrapState(queryClient)
+      .then((next) => startupInput.composingEnded().then(() => next))
+      .then((next) => {
+        if (active) setState((current) => settle(current, next))
+      })
     return () => {
       active = false
     }
-  }, [])
+  }, [queryClient, startupInput])
 
   if (state.status === "loading") {
     return <StartupScreen frame={state.frame} />
@@ -139,16 +169,15 @@ export function AppBootstrap() {
           setState(startLoading())
           void Promise.all([listWorkspaces(), loadDashboard()])
             .then(([workspaces, Dashboard]) =>
-              setState((current) =>
-                settle(current, {
-                  status: "ready",
-                  Dashboard,
-                  selection,
-                  availability: { status: "available" },
-                  workspaces,
-                })
-              )
+              afterComposing({
+                status: "ready",
+                Dashboard,
+                selection,
+                availability: { status: "available" },
+                workspaces,
+              })
             )
+            .then((next) => setState((current) => settle(current, next)))
             .catch((error: unknown) =>
               setState({ status: "error", message: messageFrom(error) })
             )
@@ -175,9 +204,9 @@ export function AppBootstrap() {
               variant="outline"
               onClick={() => {
                 setState(startLoading())
-                void fetchBootstrapState().then((next) =>
-                  setState((current) => settle(current, next))
-                )
+                void fetchBootstrapState(queryClient)
+                  .then(afterComposing)
+                  .then((next) => setState((current) => settle(current, next)))
               }}
             >
               {intl.formatMessage({
diff --git a/surfsense_local/frontend/src/features/chat/chat-composer.test.tsx b/surfsense_local/frontend/src/features/chat/chat-composer.test.tsx
index d936a6bc76..bdfafff079 100644
--- a/surfsense_local/frontend/src/features/chat/chat-composer.test.tsx
+++ b/surfsense_local/frontend/src/features/chat/chat-composer.test.tsx
@@ -5,7 +5,12 @@ import { AssistantRuntimeProvider, useLocalRuntime } from "@assistant-ui/react"
 
 import { TooltipProvider } from "@/components/ui/tooltip"
 import type { ModelSelection } from "@/features/models/selection/api"
-import type { Plugin } from "@/features/plugins/api"
+import type { Plugin, PluginTool } from "@/features/plugins/api"
+import {
+  createStartupInput,
+  type StartupInput,
+} from "@/features/startup/startup-input"
+import { StartupInputContext } from "@/features/startup/startup-input-context"
 import { render } from "@/test-utils"
 
 import { ChatComposer } from "./chat-composer"
@@ -50,32 +55,43 @@ function Harness({
   onUploadSources,
   threadMode,
   onOpenPlugins,
+  startup = null,
 }: {
   model?: ModelSelection
   onUploadSources?: (files: File[]) => void
   threadMode?: "agentic" | "basic"
   onOpenPlugins?: () => void
+  startup?: StartupInput | null
 }) {
   const runtime = useLocalRuntime({ run: async () => ({ content: [] }) })
   return (
-    <AssistantRuntimeProvider runtime={runtime}>
-      <ChatComposer
-        placement="center"
-        model={model}
-        sourceCount={0}
-        isRunning={false}
-        providerAvailable
-        onModelSetup={() => undefined}
-        onModelSelected={() => undefined}
-        readsImages={false}
-        onUploadSources={onUploadSources}
-        threadMode={threadMode}
-        onOpenPlugins={onOpenPlugins}
-      />
-    </AssistantRuntimeProvider>
+    <StartupInputContext.Provider value={startup}>
+      <AssistantRuntimeProvider runtime={runtime}>
+        <ChatComposer
+          placement="center"
+          model={model}
+          sourceCount={0}
+          isRunning={false}
+          providerAvailable
+          onModelSetup={() => undefined}
+          onModelSelected={() => undefined}
+          readsImages={false}
+          onUploadSources={onUploadSources}
+          threadMode={threadMode}
+          onOpenPlugins={onOpenPlugins}
+        />
+      </AssistantRuntimeProvider>
+    </StartupInputContext.Provider>
   )
 }
 
+// What the startup frame's box held when the dashboard took over.
+function typedDuringStartup(text: string, caret: number) {
+  const startup = createStartupInput()
+  startup.keep(text, caret)
+  return startup
+}
+
 async function openThinking(user: ReturnType<typeof userEvent.setup>) {
   await user.click(
     screen.getByRole("button", { name: "Add images, sources, and more" })
@@ -89,6 +105,110 @@ afterEach(() => {
   localStorage.clear()
 })
 
+describe("chat composer after startup", () => {
+  it("takes what was typed while the app loaded, with its caret, and keeps focus", async () => {
+    render(
+      <TooltipProvider>
+        <Harness startup={typedDuringStartup("find jobs today", 9)} />
+      </TooltipProvider>
+    )
+
+    const box = screen.getByRole<HTMLTextAreaElement>("textbox", {
+      name: "Message",
+    })
+    await waitFor(() => expect(box.value).toBe("find jobs today"))
+    expect(box.selectionStart).toBe(9)
+    expect(document.activeElement).toBe(box)
+  })
+
+  it("opens the plugin list on a mention typed while the app loaded", async () => {
+    const tool = { name: "search", enabled: true } as PluginTool
+    vi.stubGlobal(
+      "fetch",
+      vi.fn(async (input: RequestInfo | URL) =>
+        String(input) === "/plugins/third-party"
+          ? Response.json({ allowed: true })
+          : String(input) === "/plugins"
+            ? Response.json([
+                plugin({ id: "notes", name: "Notes", tools: [tool] }),
+                plugin({ id: "indeed", name: "Indeed", tools: [tool] }),
+              ])
+            : Response.json([])
+      )
+    )
+    render(
+      <TooltipProvider>
+        <Harness
+          threadMode="agentic"
+          startup={typedDuringStartup("find jobs with @ind", 19)}
+        />
+      </TooltipProvider>
+    )
+
+    const list = await screen.findByRole("listbox", { name: "Plugins" })
+    await waitFor(() => expect(list.textContent).toContain("Indeed"))
+    expect(list.textContent).not.toContain("Notes")
+  })
+
+  it("opens the plugin list on a mention the caret was left inside", async () => {
+    const tool = { name: "search", enabled: true } as PluginTool
+    vi.stubGlobal(
+      "fetch",
+      vi.fn(async (input: RequestInfo | URL) =>
+        String(input) === "/plugins/third-party"
+          ? Response.json({ allowed: true })
+          : String(input) === "/plugins"
+            ? Response.json([
+                plugin({ id: "notes", name: "Notes", tools: [tool] }),
+                plugin({ id: "indeed", name: "Indeed", tools: [tool] }),
+              ])
+            : Response.json([])
+      )
+    )
+    render(
+      <TooltipProvider>
+        <Harness
+          threadMode="agentic"
+          startup={typedDuringStartup("ask @ind about jobs", 8)}
+        />
+      </TooltipProvider>
+    )
+
+    const list = await screen.findByRole("listbox", { name: "Plugins" })
+    await waitFor(() => expect(list.textContent).toContain("Indeed"))
+    expect(
+      screen.getByRole<HTMLTextAreaElement>("textbox", { name: "Message" })
+        .selectionStart
+    ).toBe(8)
+  })
+
+  it("starts empty when another composer took the text already", async () => {
+    const startup = typedDuringStartup("find jobs", 9)
+    const first = render(
+      <TooltipProvider>
+        <Harness startup={startup} />
+      </TooltipProvider>
+    )
+    await waitFor(() =>
+      expect(
+        screen.getByRole<HTMLTextAreaElement>("textbox", { name: "Message" })
+          .value
+      ).toBe("find jobs")
+    )
+    first.unmount()
+    render(
+      <TooltipProvider>
+        <Harness startup={startup} />
+      </TooltipProvider>
+    )
+
+    expect(
+      screen.getByRole<HTMLTextAreaElement>("textbox", { name: "Message" })
+        .value
+    ).toBe("")
+  })
+})
+
 describe("chat composer", () => {
   it("stops a question at its share of the window and says so", async () => {
     vi.stubGlobal(
diff --git a/surfsense_local/frontend/src/features/chat/chat-composer.tsx b/surfsense_local/frontend/src/features/chat/chat-composer.tsx
index 1c0df78dbd..e602af2402 100644
--- a/surfsense_local/frontend/src/features/chat/chat-composer.tsx
+++ b/surfsense_local/frontend/src/features/chat/chat-composer.tsx
@@ -1,4 +1,4 @@
-import { type MouseEvent, type ReactNode } from "react"
+import { useRef, type MouseEvent, type ReactNode } from "react"
 import { ComposerPrimitive, useAuiState } from "@assistant-ui/react"
 
 import { Button } from "@/components/ui/button"
@@ -14,17 +14,19 @@ import { PluginMentions } from "@/features/plugins/in-thread/plugin-mentions"
 import { cn } from "@/lib/utils"
 import { intl } from "@/i18n/intl"
 import type { PluginsStart } from "@/features/plugins/plugins-settings"
+import { StartupDraftHandover } from "@/features/startup/startup-draft-handover"
+import { useStartupDraft } from "@/features/startup/use-startup-draft"
 import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 
 import { ComposerImage } from "./attached-image"
 import { ComposerAddMenu } from "./composer-add-menu"
+import { messageLabel, startPlaceholder } from "./composer-wording"
 import {
-  COMPOSER_CLASS,
-  COMPOSER_CONTROLS_CLASS,
-  COMPOSER_FOOTER_CLASS,
-  composerBoxClass,
-  composerInputClass,
-} from "./composer-layout"
+  ComposerBox,
+  ComposerColumn,
+  ComposerControls,
+  ComposerText,
+} from "./composer-surface"
 import { ModePicker } from "./modes/mode-picker"
 import { useNewChatMode } from "./modes/new-chat-mode"
 import { ModelPicker, modelControlButtonClassName } from "./model-picker"
@@ -218,6 +220,8 @@ export function ChatComposer({
     ({ composer }) => composer.text.length >= QUESTION_MAX_CHARS
   )
   const reveal = useStartupReveal()
+  const inputRef = useRef<HTMLTextAreaElement>(null)
+  const startupDraft = useStartupDraft()
   const dictationKeys = useDictationKeys(dictation)
   // While the microphone starts, the mic's spinner says so and nothing else moves.
   const recording = useDictationStarted(dictation)
@@ -262,7 +266,7 @@ export function ChatComposer({
     />
   )
   return (
-    <div className={COMPOSER_CLASS} data-composer-placement={placement}>
+    <ComposerColumn data-composer-placement={placement}>
       {notice ? (
         // Tucked behind the composer, which paints over its lower edge: out of
         // flow, so the composer keeps its place and its own shape.
@@ -275,60 +279,67 @@ export function ChatComposer({
           />
         </div>
       ) : null}
-      <ComposerPrimitive.Root
+      <ComposerBox
+        placement={placement}
+        render={<ComposerPrimitive.Root />}
+        {...reveal}
         className={cn(
-          composerBoxClass(placement),
           "transition-colors focus-within:border-ring/40 hover:border-ring/40",
-          reveal
+          reveal.className
         )}
       >
         <ComposerPrimitive.Unstable_TriggerPopoverRoot>
           {/* Plugins reach Agentic threads only, so only their composer names tools. */}
           {agentic ? <PluginMentions /> : null}
+          {startupDraft ? (
+            <StartupDraftHandover draft={startupDraft} inputRef={inputRef} />
+          ) : null}
           {placement === "bottom" ? addMenu("-mr-1.5 mb-0.5") : null}
-          <ComposerPrimitive.Input
-            autoFocus
-            unstable_focusOnThreadSwitched
-            aria-describedby={describedBy}
-            // Held whenever a send could not go anywhere: no model, a model that
-            // can't be used, or egress to it still off.
-            disabled={inputDisabled}
-            // Voice input writes the text meanwhile; the box keeps its focus and keys.
-            readOnly={dictationKeys.recording}
-            cancelOnEscape={!dictationKeys.recording}
-            onKeyDown={dictationKeys.onKeyDown}
-            className={composerInputClass(placement)}
-            placeholder={
-              blockedPlaceholder !== undefined
-                ? blockedPlaceholder
-                : !model || providerAvailable
-                  ? placement === "center"
-                    ? intl.formatMessage({
-                        id: "chat_composer_start_placeholder",
-                        defaultMessage: "Turn your sources into answers",
-                      })
-                    : intl.formatMessage({
-                        id: "chat_composer_follow_up_placeholder",
-                        defaultMessage: "Follow up on this answer",
-                      })
-                  : intl.formatMessage({
-                      id: "chat_composer_provider_offline_placeholder",
-                      defaultMessage: "Reconnect your model provider to send",
-                    })
+          {/* assistant-ui's own input, autosizing, in the composer's shape. */}
+          <ComposerText
+            placement={placement}
+            render={
+              <ComposerPrimitive.Input
+                ref={inputRef}
+                // The handover focuses it with the caret where the user left it.
+                autoFocus={startupDraft === null}
+                unstable_focusOnThreadSwitched
+                aria-describedby={describedBy}
+                // Held whenever a send could not go anywhere: no model, a model that
+                // can't be used, or egress to it still off.
+                disabled={inputDisabled}
+                // Voice input writes the text meanwhile; the box keeps its focus and keys.
+                readOnly={dictationKeys.recording}
+                cancelOnEscape={!dictationKeys.recording}
+                onKeyDown={dictationKeys.onKeyDown}
+                placeholder={
+                  blockedPlaceholder !== undefined
+                    ? blockedPlaceholder
+                    : !model || providerAvailable
+                      ? placement === "center"
+                        ? startPlaceholder()
+                        : intl.formatMessage({
+                            id: "chat_composer_follow_up_placeholder",
+                            defaultMessage: "Follow up on this answer",
+                          })
+                      : intl.formatMessage({
+                          id: "chat_composer_provider_offline_placeholder",
+                          defaultMessage:
+                            "Reconnect your model provider to send",
+                        })
+                }
+                submitMode="enter"
+                addAttachmentOnPaste={readsImages && !agentic}
+                rows={1}
+                maxLength={QUESTION_MAX_CHARS}
+                aria-label={messageLabel()}
+              />
             }
-            submitMode="enter"
-            addAttachmentOnPaste={readsImages && !agentic}
-            rows={1}
-            maxLength={QUESTION_MAX_CHARS}
-            aria-label={intl.formatMessage({
-              id: "chat_composer_message_aria",
-              defaultMessage: "Message",
-            })}
           />
           {placement === "center" ? (
             // Its own row under the text, never over it: long text scrolls
             // inside the box and stops above the controls.
-            <div className={COMPOSER_CONTROLS_CLASS}>
+            <ComposerControls>
               <div className="flex shrink-0 items-center gap-1">
                 {addMenu("")}
                 {model ? <ModePicker model={model} threadMode={null} /> : null}
@@ -353,7 +364,7 @@ export function ChatComposer({
                   onStopRecording={stopRecording}
                 />
               </div>
-            </div>
+            </ComposerControls>
           ) : (
             <>
               {recording ? (
@@ -371,7 +382,7 @@ export function ChatComposer({
             </>
           )}
         </ComposerPrimitive.Unstable_TriggerPopoverRoot>
-      </ComposerPrimitive.Root>
+      </ComposerBox>
       {placement === "center" && limitNotice ? (
         <p
           role="status"
@@ -381,7 +392,13 @@ export function ChatComposer({
         </p>
       ) : null}
       {placement === "bottom" ? (
-        <div className={cn(COMPOSER_FOOTER_CLASS, reveal)}>
+        <div
+          {...reveal}
+          className={cn(
+            "mt-1 flex min-h-7 items-center justify-between gap-3 px-2",
+            reveal.className
+          )}
+        >
           {/* Keeps its line and leaves the mode and the model name what is
           left, never less than 13rem, so a long translation truncates the
           name before wrapping. */}
@@ -418,6 +435,6 @@ export function ChatComposer({
           </div>
         </div>
       ) : null}
-    </div>
+    </ComposerColumn>
   )
 }
diff --git a/surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts b/surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts
deleted file mode 100644
index 8c49f79ba5..0000000000
--- a/surfsense_local/frontend/src/features/chat/chat-viewport-layout.ts
+++ /dev/null
@@ -1,19 +0,0 @@
-// The conversation's scroll area and its sticky composer footer, shared by
-// ChatViewport and the startup frame, which draws them without the runtime.
-
-export const THREAD_BODY_CLASS = "relative flex min-h-0 flex-1 flex-col"
-
-export const VIEWPORT_BOX_CLASS =
-  "relative flex min-h-0 w-full min-w-0 flex-1 flex-col"
-
-// Top edge only: a bottom mask would fade the sticky composer.
-export const VIEWPORT_CLASS =
-  "flex min-h-0 w-full min-w-0 flex-1 scroll-fade-t flex-col overflow-y-auto px-4 [--scroll-fade-reveal:24px] scroll-fade-6"
-
-// The gutter moves the centred column on systems with classic scrollbars.
-export const VIEWPORT_STYLE = { scrollbarGutter: "stable" } as const
-
-export const VIEWPORT_FOOTER_CLASS =
-  "sticky bottom-0 z-20 -mx-4 mt-auto shrink-0 bg-gradient-to-t from-background via-background to-transparent px-4 pb-1"
-
-export const VIEWPORT_FOOTER_INNER_CLASS = "relative mx-auto w-full max-w-3xl"
diff --git a/surfsense_local/frontend/src/features/chat/chat-viewport.tsx b/surfsense_local/frontend/src/features/chat/chat-viewport.tsx
index a91a3f6c2d..06652f5dc9 100644
--- a/surfsense_local/frontend/src/features/chat/chat-viewport.tsx
+++ b/surfsense_local/frontend/src/features/chat/chat-viewport.tsx
@@ -5,13 +5,7 @@ import { ArrowDownIcon } from "@/components/ui/icons"
 import { intl } from "@/i18n/intl"
 import { cn } from "@/lib/utils"
 
-import {
-  VIEWPORT_BOX_CLASS,
-  VIEWPORT_CLASS,
-  VIEWPORT_FOOTER_CLASS,
-  VIEWPORT_FOOTER_INNER_CLASS,
-  VIEWPORT_STYLE,
-} from "./chat-viewport-layout"
+import { ConversationScroll } from "./conversation-scroll"
 
 // Raised clear of a notice tucked behind the composer's top edge, whose visible
 // part is about 28px tall.
@@ -47,27 +41,27 @@ export function ChatViewport({
   footerHasNotice?: boolean
 }) {
   return (
-    <div className={VIEWPORT_BOX_CLASS}>
-      <ThreadPrimitive.Viewport
-        turnAnchor="top"
-        className={VIEWPORT_CLASS}
-        style={VIEWPORT_STYLE}
-        autoScroll
-        scrollToBottomOnRunStart
-        scrollToBottomOnInitialize
-        scrollToBottomOnThreadSwitch
-        data-chat-viewport
-      >
-        {children}
-        {footer ? (
-          <ThreadPrimitive.ViewportFooter className={VIEWPORT_FOOTER_CLASS}>
-            <div className={VIEWPORT_FOOTER_INNER_CLASS}>
-              <ScrollToBottom raised={footerHasNotice} />
-              {footer}
-            </div>
-          </ThreadPrimitive.ViewportFooter>
-        ) : null}
-      </ThreadPrimitive.Viewport>
-    </div>
+    <ConversationScroll
+      render={
+        <ThreadPrimitive.Viewport
+          turnAnchor="top"
+          autoScroll
+          scrollToBottomOnRunStart
+          scrollToBottomOnInitialize
+          scrollToBottomOnThreadSwitch
+          data-chat-viewport
+        />
+      }
+    >
+      {children}
+      {footer ? (
+        <ThreadPrimitive.ViewportFooter className="sticky bottom-0 z-20 -mx-4 mt-auto shrink-0 bg-gradient-to-t from-background via-background to-transparent px-4 pb-1">
+          <div className="relative mx-auto w-full max-w-3xl">
+            <ScrollToBottom raised={footerHasNotice} />
+            {footer}
+          </div>
+        </ThreadPrimitive.ViewportFooter>
+      ) : null}
+    </ConversationScroll>
   )
 }
diff --git a/surfsense_local/frontend/src/features/chat/composer-layout.ts b/surfsense_local/frontend/src/features/chat/composer-layout.ts
deleted file mode 100644
index f6faf262ce..0000000000
--- a/surfsense_local/frontend/src/features/chat/composer-layout.ts
+++ /dev/null
@@ -1,33 +0,0 @@
-import { cn } from "@/lib/utils"
-
-// The composer's geometry, shared by ChatComposer and the startup frame's
-// outline of it, so the box the frame draws is the box that replaces it.
-
-export type ComposerPlacement = "center" | "bottom"
-
-export const COMPOSER_CLASS = "relative mx-auto w-full max-w-2xl"
-
-// No top padding: the text scrolls up to the border, not to a strip above
-// it, and holds that space in its own padding instead.
-export function composerBoxClass(placement: ComposerPlacement) {
-  return cn(
-    "relative rounded-2xl border bg-card px-1.5 pb-1.5 shadow-sm",
-    placement === "bottom" && "flex items-end gap-2"
-  )
-}
-
-// One line tall at rest: an outline holds a line of nothing to match it.
-export function composerInputClass(placement: ComposerPlacement) {
-  return cn(
-    "max-h-44 resize-none bg-transparent px-2 pt-4 pb-2.5 text-sm outline-none placeholder:text-muted-foreground",
-    placement === "center" ? "block min-h-12 w-full" : "min-h-10 flex-1"
-  )
-}
-
-// At least as tall as its icon-lg buttons, so the outline, which has none,
-// keeps the height they give it.
-export const COMPOSER_CONTROLS_CLASS =
-  "flex min-h-9 items-center justify-between gap-2"
-
-export const COMPOSER_FOOTER_CLASS =
-  "mt-1 flex min-h-7 items-center justify-between gap-3 px-2"
diff --git a/surfsense_local/frontend/src/features/chat/composer-outline.tsx b/surfsense_local/frontend/src/features/chat/composer-outline.tsx
deleted file mode 100644
index 68c92eca96..0000000000
--- a/surfsense_local/frontend/src/features/chat/composer-outline.tsx
+++ /dev/null
@@ -1,36 +0,0 @@
-import {
-  COMPOSER_CLASS,
-  COMPOSER_CONTROLS_CLASS,
-  COMPOSER_FOOTER_CLASS,
-  composerBoxClass,
-  composerInputClass,
-  type ComposerPlacement,
-} from "./composer-layout"
-
-/**
- * The composer's empty box, the size ChatComposer draws it, for the startup
- * frame: its text and controls fade in when the real one takes its place.
- */
-export function ComposerOutline({
-  placement,
-}: {
-  placement: ComposerPlacement
-}) {
-  return (
-    <div
-      aria-hidden="true"
-      className={COMPOSER_CLASS}
-      data-composer-outline={placement}
-    >
-      <div className={composerBoxClass(placement)}>
-        <div className={composerInputClass(placement)}>&nbsp;</div>
-        {placement === "center" ? (
-          <div className={COMPOSER_CONTROLS_CLASS} />
-        ) : null}
-      </div>
-      {placement === "bottom" ? (
-        <div className={COMPOSER_FOOTER_CLASS} />
-      ) : null}
-    </div>
-  )
-}
diff --git a/surfsense_local/frontend/src/features/chat/composer-surface.tsx b/surfsense_local/frontend/src/features/chat/composer-surface.tsx
new file mode 100644
index 0000000000..7303a7af30
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/composer-surface.tsx
@@ -0,0 +1,101 @@
+import { mergeProps } from "@base-ui/react/merge-props"
+import { useRender } from "@base-ui/react/use-render"
+import type { ComponentProps } from "react"
+import TextareaAutosize from "react-textarea-autosize"
+
+import { cn } from "@/lib/utils"
+
+// The composer's shape, drawn by ChatComposer around assistant-ui's parts and
+// by the startup frame on its own: the box it shows first is the box that
+// stays. A new chat's composer is "center", a thread's "bottom".
+
+type Placement = "center" | "bottom"
+
+/** The column the composer sits in, with what goes above and below its box. */
+export function ComposerColumn({ className, ...props }: ComponentProps<"div">) {
+  return (
+    <div
+      className={cn("relative mx-auto w-full max-w-2xl", className)}
+      {...props}
+    />
+  )
+}
+
+/** The bordered box. ChatComposer renders it as the composer's form. */
+export function ComposerBox({
+  placement,
+  className,
+  render,
+  ...props
+}: useRender.ComponentProps<"div"> & { placement: Placement }) {
+  return useRender({
+    defaultTagName: "div",
+    render,
+    props: mergeProps<"div">(
+      {
+        // No top padding: text scrolls up to the border, not to a strip above
+        // it, and holds that space in its own padding instead.
+        className: cn(
+          "relative rounded-2xl border bg-card px-1.5 pb-1.5 shadow-sm",
+          placement === "bottom" && "flex items-end gap-2",
+          className
+        ),
+      },
+      props
+    ),
+  })
+}
+
+/**
+ * The text box: one line at rest, growing with its text. ChatComposer renders
+ * it as assistant-ui's input; on its own it is the autosizing textarea that
+ * input is built on, so the startup frame's box grows exactly as the
+ * composer's does, in every browser. It stays solid while the controls
+ * around it fade in at startup.
+ */
+export function ComposerText({
+  placement,
+  className,
+  render,
+  ...props
+}: useRender.ComponentProps<"textarea"> & { placement: Placement }) {
+  return useRender({
+    defaultTagName: "textarea",
+    render: render ?? <TextareaAutosize />,
+    // Marked for the startup fade, which leaves it alone.
+    state: { revealSteady: true },
+    stateAttributesMapping: {
+      revealSteady: () => ({ "data-reveal-steady": "" }),
+    },
+    props: mergeProps<"textarea">(
+      {
+        rows: 1,
+        className: cn(
+          "max-h-44 resize-none bg-transparent px-2 pt-4 pb-2.5 text-sm outline-none placeholder:text-muted-foreground",
+          placement === "center" ? "block min-h-12 w-full" : "min-h-10 flex-1",
+          className
+        ),
+      },
+      props
+    ),
+  })
+}
+
+/**
+ * A new chat's row of controls under its text. At least as tall as its
+ * icon-lg buttons, so the startup frame's, which has none, matches.
+ */
+export function ComposerControls({
+  className,
+  ...props
+}: ComponentProps<"div">) {
+  return (
+    <div
+      className={cn(
+        "flex min-h-9 items-center justify-between gap-2",
+        className
+      )}
+      {...props}
+    />
+  )
+}
diff --git a/surfsense_local/frontend/src/features/chat/composer-wording.ts b/surfsense_local/frontend/src/features/chat/composer-wording.ts
new file mode 100644
index 0000000000..76ce297869
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/composer-wording.ts
@@ -0,0 +1,17 @@
+import { intl } from "@/i18n/intl"
+
+// Shared with the startup frame's box, so it reads as the composer it becomes.
+
+export function startPlaceholder() {
+  return intl.formatMessage({
+    id: "chat_composer_start_placeholder",
+    defaultMessage: "Turn your sources into answers",
+  })
+}
+
+export function messageLabel() {
+  return intl.formatMessage({
+    id: "chat_composer_message_aria",
+    defaultMessage: "Message",
+  })
+}
diff --git a/surfsense_local/frontend/src/features/chat/conversation-column.tsx b/surfsense_local/frontend/src/features/chat/conversation-column.tsx
index 4693694d92..90d25d3a35 100644
--- a/surfsense_local/frontend/src/features/chat/conversation-column.tsx
+++ b/surfsense_local/frontend/src/features/chat/conversation-column.tsx
@@ -1,4 +1,4 @@
-import type { ReactNode } from "react"
+import type { ComponentProps, ReactNode } from "react"
 
 import { ColumnHeader } from "@/components/ui/column-header"
 import { intl } from "@/i18n/intl"
@@ -6,17 +6,17 @@ import { cn } from "@/lib/utils"
 
 /**
  * The conversation's box and its header line, which the startup frame draws
- * empty. `heading` precedes the header; `headerClassName` styles its line.
+ * empty. `heading` precedes the header; `headerProps` go on its line.
  */
 export function ConversationColumn({
   heading,
   header,
-  headerClassName,
+  headerProps,
   children,
 }: {
   heading?: ReactNode
   header?: ReactNode
-  headerClassName?: string
+  headerProps?: ComponentProps<"div">
   children: ReactNode
 }) {
   return (
@@ -29,7 +29,10 @@ export function ConversationColumn({
     >
       {heading}
       <header>
-        <ColumnHeader className={cn("px-5", headerClassName)}>
+        <ColumnHeader
+          {...headerProps}
+          className={cn("px-5", headerProps?.className)}
+        >
           {header}
         </ColumnHeader>
       </header>
diff --git a/surfsense_local/frontend/src/features/chat/conversation-scroll.tsx b/surfsense_local/frontend/src/features/chat/conversation-scroll.tsx
new file mode 100644
index 0000000000..6ea94da68f
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/conversation-scroll.tsx
@@ -0,0 +1,55 @@
+import { mergeProps } from "@base-ui/react/merge-props"
+import { useRender } from "@base-ui/react/use-render"
+
+import { cn } from "@/lib/utils"
+
+// The conversation's body and scroll area, drawn by the thread panel around
+// assistant-ui's parts and by the startup frame on its own.
+
+/** Under the conversation's header, the rest of its height. */
+export function ConversationBody({
+  className,
+  render,
+  ...props
+}: useRender.ComponentProps<"div">) {
+  return useRender({
+    defaultTagName: "div",
+    render,
+    props: mergeProps<"div">(
+      { className: cn("relative flex min-h-0 flex-1 flex-col", className) },
+      props
+    ),
+  })
+}
+
+/**
+ * The scrolling column of messages, or a new chat's centred box. Its gutter
+ * stays, so the centred column sits in one place with or without a scrollbar.
+ */
+export function ConversationScroll({
+  className,
+  style,
+  render,
+  ...props
+}: useRender.ComponentProps<"div">) {
+  const scroll = useRender({
+    defaultTagName: "div",
+    render,
+    props: mergeProps<"div">(
+      {
+        // Top edge only: a bottom mask would fade the sticky composer.
+        className: cn(
+          "flex min-h-0 w-full min-w-0 flex-1 scroll-fade-t flex-col overflow-y-auto px-4 [--scroll-fade-reveal:24px] scroll-fade-6",
+          className
+        ),
+        style: { scrollbarGutter: "stable", ...style },
+      },
+      props
+    ),
+  })
+  return (
+    <div className="relative flex min-h-0 w-full min-w-0 flex-1 flex-col">
+      {scroll}
+    </div>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode-store.ts b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode-store.ts
new file mode 100644
index 0000000000..8d27f9d3ec
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode-store.ts
@@ -0,0 +1,22 @@
+import type { ChatMode } from "@/features/models/capability/api"
+
+// Read by the startup frame, so this file imports nothing at run time.
+const NEXT_NEW_CHAT_MODE_KEY = "surfsense:next-new-chat-mode:v1"
+
+/** The mode a new chat opens in at the next launch, as last worked out. */
+export function readNextNewChatMode(): ChatMode | null {
+  try {
+    const stored = localStorage.getItem(NEXT_NEW_CHAT_MODE_KEY)
+    return stored === "agentic" || stored === "basic" ? stored : null
+  } catch {
+    return null
+  }
+}
+
+export function writeNextNewChatMode(mode: ChatMode): void {
+  try {
+    localStorage.setItem(NEXT_NEW_CHAT_MODE_KEY, mode)
+  } catch {
+    // The frame then offers no mentions; the composer still does.
+  }
+}
diff --git a/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.test.ts b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.test.ts
new file mode 100644
index 0000000000..cb5e840cff
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.test.ts
@@ -0,0 +1,57 @@
+import { afterEach, describe, expect, it } from "vitest"
+
+import type {
+  ChatMode,
+  ModelCapability,
+} from "@/features/models/capability/api"
+import type { ModelSelection } from "@/features/models/selection/api"
+import { readStartupLayout } from "@/features/startup/startup-layout"
+
+import { rememberNextNewChatMode } from "./next-new-chat-mode"
+
+function model(defaultMode: ChatMode, agenticAllowed = true): ModelSelection {
+  return {
+    model_type: "text_gen",
+    provider: "openai_compatible",
+    connection_id: 1,
+    name: "anthropic/claude-fable-5",
+    updated_at: "2026-11-01T00:00:00Z",
+    capability: {
+      modes: {
+        agentic_allowed: agenticAllowed,
+        blocked: null,
+        default_mode: defaultMode,
+        reason: { code: "assumed", values: {} },
+      },
+    } as unknown as ModelCapability,
+  }
+}
+
+afterEach(() => localStorage.clear())
+
+describe("the next new chat's mode, for the startup frame", () => {
+  it("offers mentions when the model opens new chats in Agentic", () => {
+    rememberNextNewChatMode(model("agentic"))
+    expect(readStartupLayout().mentions).toBe(true)
+  })
+
+  it("offers none when the model opens new chats in Q&A", () => {
+    rememberNextNewChatMode(model("basic"))
+    expect(readStartupLayout().mentions).toBe(false)
+  })
+
+  it("offers none when the model may not run Agentic chats", () => {
+    rememberNextNewChatMode(model("agentic", false))
+    expect(readStartupLayout().mentions).toBe(false)
+  })
+
+  it("offers none without a model", () => {
+    rememberNextNewChatMode(model("agentic"))
+    rememberNextNewChatMode(null)
+    expect(readStartupLayout().mentions).toBe(false)
+  })
+
+  it("offers none before anything was remembered", () => {
+    expect(readStartupLayout().mentions).toBe(false)
+  })
+})
diff --git a/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts
new file mode 100644
index 0000000000..39fa3c043e
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts
@@ -0,0 +1,13 @@
+import type { ModelSelection } from "@/features/models/selection/api"
+
+import { resolveNewChatChoice } from "./new-chat-mode"
+import { writeNextNewChatMode } from "./next-new-chat-mode-store"
+
+/**
+ * Remembers the mode a new chat on this model opens in after a restart: its
+ * default, since a mode picked this session is not kept. The startup frame
+ * reads it to offer plugin mentions before the model has loaded.
+ */
+export function rememberNextNewChatMode(model: ModelSelection | null): void {
+  writeNextNewChatMode(resolveNewChatChoice(model, {})?.mode ?? "basic")
+}
diff --git a/surfsense_local/frontend/src/features/chat/thread-panel.tsx b/surfsense_local/frontend/src/features/chat/thread-panel.tsx
index 073027205f..1a79f0bc7d 100644
--- a/surfsense_local/frontend/src/features/chat/thread-panel.tsx
+++ b/surfsense_local/frontend/src/features/chat/thread-panel.tsx
@@ -32,8 +32,8 @@ import { useStartupReveal } from "@/features/startup/use-startup-reveal"
 import type { ChatThread } from "./api"
 import { ChatComposer } from "./chat-composer"
 import { ChatViewport } from "./chat-viewport"
-import { THREAD_BODY_CLASS } from "./chat-viewport-layout"
 import { ConversationColumn } from "./conversation-column"
+import { ConversationBody } from "./conversation-scroll"
 import { composerAttachmentOf } from "./image-attachments"
 import { LiveThreadRuntime, type LiveThreadSource } from "./live-thread-runtime"
 import { AssistantMessage, UserMessage } from "./message"
@@ -301,7 +301,7 @@ export function ThreadPanel({
             </h2>
           )
         }
-        headerClassName={reveal}
+        headerProps={reveal}
         header={
           <>
             {thread == null ? null : editing ? (
@@ -420,7 +420,7 @@ export function ThreadPanel({
           </>
         }
       >
-        <ThreadPrimitive.Root className={THREAD_BODY_CLASS}>
+        <ConversationBody render={<ThreadPrimitive.Root />}>
           <ChatViewport footer={bottomFooter} footerHasNotice={notice != null}>
             {isLoading ? <MessagesSkeleton /> : null}
 
@@ -434,7 +434,7 @@ export function ThreadPanel({
               </ThreadPrimitive.Messages>
             ) : null}
           </ChatViewport>
-        </ThreadPrimitive.Root>
+        </ConversationBody>
       </ConversationColumn>
     </LiveThreadRuntime>
   )
diff --git a/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx b/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx
index e483c35b44..1460579967 100644
--- a/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/dashboard-page.tsx
@@ -25,6 +25,7 @@ import {
   useNewChatChoice,
 } from "@/features/chat/modes/new-chat-mode"
 import { NewChatModeProvider } from "@/features/chat/modes/new-chat-mode-provider"
+import { rememberNextNewChatMode } from "@/features/chat/modes/next-new-chat-mode"
 import { canSkipThinking } from "@/features/chat/thinking-preference"
 import { ThreadPanel } from "@/features/chat/thread-panel"
 import { useChatRuntime } from "@/features/chat/use-chat-runtime"
@@ -395,7 +396,10 @@ function WorkspaceDashboard({
     <OfficeOfferContext.Provider value={officeOffer}>
       <OpenArtifactContext.Provider value={openArtifact}>
         <div className="titlebar-controls">
-          <div className={cn("titlebar-controls-end", reveal)}>
+          <div
+            {...reveal}
+            className={cn("titlebar-controls-end", reveal.className)}
+          >
             <TitleBarUpdate />
             {/* Docker's browser has no title bar: there the toggle sits on a heading's line. */}
             {IN_CONTAINER
@@ -706,6 +710,9 @@ export function DashboardPage({
     Boolean(workspaces.activeWorkspace)
   )
 
+  // For the startup frame, which offers plugin mentions before this loads.
+  useEffect(() => rememberNextNewChatMode(selection), [selection])
+
   // Checked when the model changes, when settings close (a key entered again,
   // egress switched, a connection edited), after egress is allowed, and after a
   // send was refused for Claude Code.
diff --git a/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx b/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
index 440bdb0744..cbc0db4832 100644
--- a/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
@@ -176,7 +176,10 @@ export const LeftSidebar = memo(function LeftSidebar({
   return (
     <SidebarFrame
       nav={
-        <div className={cn("flex flex-col px-3 pt-2", reveal)}>
+        <div
+          {...reveal}
+          className={cn("flex flex-col px-3 pt-2", reveal.className)}
+        >
           <SidebarNavButton
             label={intl.formatMessage({
               id: "dashboard_sidebar_new_chat_button",
@@ -201,14 +204,17 @@ export const LeftSidebar = memo(function LeftSidebar({
       }
     >
       <div
+        {...reveal}
         className={cn(
           "flex min-h-0 min-w-0 flex-1 flex-col overflow-hidden px-3 pt-5 pb-2",
-          reveal
+          reveal.className
         )}
       >
         {sources}
       </div>
-      <div className={cn("contents", reveal)}>{footer}</div>
+      <div {...reveal} className={cn("contents", reveal.className)}>
+        {footer}
+      </div>
       <ChatsDialog
         open={chatsOpen}
         onOpenChange={setChatsOpen}
diff --git a/surfsense_local/frontend/src/features/dashboard/right-panel.tsx b/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
index 3f2a1e2cf7..f810a3cb52 100644
--- a/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
@@ -43,12 +43,15 @@ export function RightPanel({
           </h2>
           {toggle}
         </ColumnHeader>
-        <div className={cn("px-3 pt-2", reveal)}>{studio}</div>
+        <div {...reveal} className={cn("px-3 pt-2", reveal.className)}>
+          {studio}
+        </div>
       </header>
       <div
+        {...reveal}
         className={cn(
           "min-h-0 min-w-0 flex-1 overflow-hidden px-3 pb-2",
-          reveal
+          reveal.className
         )}
       >
         {artifacts}
diff --git a/surfsense_local/frontend/src/features/plugins/in-thread/mention-list.tsx b/surfsense_local/frontend/src/features/plugins/in-thread/mention-list.tsx
new file mode 100644
index 0000000000..aa6af89652
--- /dev/null
+++ b/surfsense_local/frontend/src/features/plugins/in-thread/mention-list.tsx
@@ -0,0 +1,68 @@
+import { mergeProps } from "@base-ui/react/merge-props"
+import { useRender } from "@base-ui/react/use-render"
+import type { ComponentProps } from "react"
+
+import { intl } from "@/i18n/intl"
+import { cn } from "@/lib/utils"
+
+// The `@` list's shape, drawn by PluginMentions around assistant-ui's popover
+// and by the startup frame as a skeleton, so the list opens where the
+// skeleton stood.
+
+/** Above the composer's text, the width of its box. Named "Plugins". */
+export function MentionList({
+  className,
+  render,
+  ...props
+}: useRender.ComponentProps<"div">) {
+  return useRender({
+    defaultTagName: "div",
+    render,
+    props: mergeProps<"div">(
+      {
+        "aria-label": intl.formatMessage({
+          id: "plugins_mentions_aria",
+          defaultMessage: "Plugins",
+        }),
+        className: cn(
+          "absolute inset-x-1.5 bottom-full z-50 mb-2 max-h-64 overflow-y-auto rounded-xl border bg-popover p-1 text-popover-foreground shadow-md",
+          className
+        ),
+      },
+      props
+    ),
+  })
+}
+
+/** A plugin's row: its name, then a line about it. */
+export function MentionItem({
+  className,
+  render,
+  ...props
+}: useRender.ComponentProps<"div">) {
+  return useRender({
+    defaultTagName: "div",
+    render,
+    props: mergeProps<"div">(
+      {
+        className: cn(
+          "flex w-full flex-col items-start gap-0.5 rounded-lg px-2 py-1.5 text-left text-sm outline-none",
+          className
+        ),
+      },
+      props
+    ),
+  })
+}
+
+export function MentionItemDescription({
+  className,
+  ...props
+}: ComponentProps<"span">) {
+  return (
+    <span
+      className={cn("line-clamp-1 text-xs text-muted-foreground", className)}
+      {...props}
+    />
+  )
+}
diff --git a/surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.test.ts b/surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.test.ts
new file mode 100644
index 0000000000..4cc0e3eabe
--- /dev/null
+++ b/surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.test.ts
@@ -0,0 +1,26 @@
+import { describe, expect, it } from "vitest"
+
+import { MENTION_TRIGGER, mentionAt } from "./mention-matcher"
+
+describe("the mention at the caret", () => {
+  it.each([
+    ["@", 1, { query: "", offset: 0, endOffset: 1 }],
+    ["check @", 7, { query: "", offset: 6, endOffset: 7 }],
+    ["find jobs with @ind", 19, { query: "ind", offset: 15, endOffset: 19 }],
+    ["ask @goo about it", 8, { query: "goo", offset: 4, endOffset: 8 }],
+    ["ask @goo about it", 6, { query: "g", offset: 4, endOffset: 6 }],
+    ["line one\n@not", 13, { query: "not", offset: 9, endOffset: 13 }],
+  ])("is found in %j at %i", (text, caret, match) => {
+    expect(mentionAt(text, MENTION_TRIGGER, caret)).toEqual(match)
+  })
+
+  it.each([
+    ["no mention here", 15],
+    ["ask @goo about it", 17],
+    ["mail me@ind.com", 15],
+    ["@goo ", 5],
+    ["", 0],
+  ])("is not found in %j at %i", (text, caret) => {
+    expect(mentionAt(text, MENTION_TRIGGER, caret)).toBeNull()
+  })
+})
diff --git a/surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.ts b/surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.ts
new file mode 100644
index 0000000000..c6a98ad229
--- /dev/null
+++ b/surfsense_local/frontend/src/features/plugins/in-thread/mention-matcher.ts
@@ -0,0 +1,30 @@
+import type {
+  Unstable_TriggerMatch,
+  Unstable_TriggerMatcher,
+} from "@assistant-ui/react"
+
+export const MENTION_TRIGGER = "@"
+
+function isSpace(char: string) {
+  return char.trim() === ""
+}
+
+/**
+ * The plugin mention the caret is in: a word that starts with the trigger, up
+ * to the caret. The composer's `@` list and the startup frame both use it,
+ * so a word opens the list in one exactly when it does in the other.
+ */
+export const mentionAt: Unstable_TriggerMatcher = (
+  text,
+  trigger,
+  caret
+): Unstable_TriggerMatch | null => {
+  let start = caret
+  while (start > 0 && !isSpace(text.charAt(start - 1))) start -= 1
+  if (!text.startsWith(trigger, start)) return null
+  return {
+    query: text.slice(start + trigger.length, caret),
+    offset: start,
+    endOffset: caret,
+  }
+}
diff --git a/surfsense_local/frontend/src/features/plugins/in-thread/mention-queries.ts b/surfsense_local/frontend/src/features/plugins/in-thread/mention-queries.ts
new file mode 100644
index 0000000000..7f99132179
--- /dev/null
+++ b/surfsense_local/frontend/src/features/plugins/in-thread/mention-queries.ts
@@ -0,0 +1,21 @@
+import { queryOptions } from "@tanstack/react-query"
+
+import {
+  listPlugins,
+  pluginsQueryKey,
+  thirdPartyAllowed,
+  thirdPartyQueryKey,
+} from "../api"
+
+// What the `@` list reads, shared with startup, which loads it early so the
+// list is ready when the composer arrives.
+
+export const pluginListQuery = queryOptions({
+  queryKey: pluginsQueryKey,
+  queryFn: ({ signal }) => listPlugins(signal),
+})
+
+export const thirdPartyQuery = queryOptions({
+  queryKey: thirdPartyQueryKey,
+  queryFn: ({ signal }) => thirdPartyAllowed(signal),
+})
diff --git a/surfsense_local/frontend/src/features/plugins/in-thread/plugin-mentions.tsx b/surfsense_local/frontend/src/features/plugins/in-thread/plugin-mentions.tsx
index 33df53229c..ffbfed2f2f 100644
--- a/surfsense_local/frontend/src/features/plugins/in-thread/plugin-mentions.tsx
+++ b/surfsense_local/frontend/src/features/plugins/in-thread/plugin-mentions.tsx
@@ -9,15 +9,17 @@ import { useMemo } from "react"
 
 import { intl } from "@/i18n/intl"
 
-import {
-  listPlugins,
-  pluginsQueryKey,
-  thirdPartyAllowed,
-  thirdPartyQueryKey,
-  type Plugin,
-} from "../api"
+import type { Plugin } from "../api"
 import { licenseRefused } from "../license-refused"
 
+import {
+  MentionItem,
+  MentionItemDescription,
+  MentionList,
+} from "./mention-list"
+import { MENTION_TRIGGER, mentionAt } from "./mention-matcher"
+import { pluginListQuery, thirdPartyQuery } from "./mention-queries"
+
 // A plugin id: lowercase words joined by hyphens.
 const MENTION = /@([a-z0-9-]+)/g
 
@@ -109,14 +111,8 @@ function emptyMessage(plugins: Plugin[], thirdPartyOn: boolean) {
 
 /** Typing `@` in an Agentic composer lists the ready plugins to name for the agent, which picks their tools. */
 export function PluginMentions() {
-  const plugins = useQuery({
-    queryKey: pluginsQueryKey,
-    queryFn: ({ signal }) => listPlugins(signal),
-  })
-  const thirdParty = useQuery({
-    queryKey: thirdPartyQueryKey,
-    queryFn: ({ signal }) => thirdPartyAllowed(signal),
-  })
+  const plugins = useQuery(pluginListQuery)
+  const thirdParty = useQuery(thirdPartyQuery)
   const items = useMemo(
     () => readyMentions(plugins.data ?? [], thirdParty.data ?? false),
     [plugins.data, thirdParty.data]
@@ -129,14 +125,15 @@ export function PluginMentions() {
   })
 
   return (
-    <ComposerPrimitive.Unstable_TriggerPopover
-      char="@"
-      adapter={mention.adapter}
-      aria-label={intl.formatMessage({
-        id: "plugins_mentions_aria",
-        defaultMessage: "Plugins",
-      })}
-      className="absolute inset-x-1.5 bottom-full z-50 mb-2 max-h-64 overflow-y-auto rounded-xl border bg-popover p-1 text-popover-foreground shadow-md"
+    <MentionList
+      render={
+        <ComposerPrimitive.Unstable_TriggerPopover
+          char={MENTION_TRIGGER}
+          adapter={mention.adapter}
+          // The rule the startup frame's box uses too.
+          matcher={mentionAt}
+        />
+      }
     >
       <ComposerPrimitive.Unstable_TriggerPopover.Directive
         formatter={mention.directive.formatter}
@@ -154,23 +151,27 @@ export function PluginMentions() {
             </p>
           ) : (
             found.map((item, index) => (
-              <ComposerPrimitive.Unstable_TriggerPopoverItem
+              <MentionItem
                 key={item.id}
-                item={item}
-                index={index}
-                className="flex w-full flex-col items-start gap-0.5 rounded-lg px-2 py-1.5 text-left text-sm outline-none hover:bg-accent data-[highlighted]:bg-accent"
+                render={
+                  <ComposerPrimitive.Unstable_TriggerPopoverItem
+                    item={item}
+                    index={index}
+                  />
+                }
+                className="hover:bg-accent data-[highlighted]:bg-accent"
               >
                 <span>{item.label}</span>
                 {item.description ? (
-                  <span className="line-clamp-1 text-xs text-muted-foreground">
+                  <MentionItemDescription>
                     {item.description}
-                  </span>
+                  </MentionItemDescription>
                 ) : null}
-              </ComposerPrimitive.Unstable_TriggerPopoverItem>
+              </MentionItem>
             ))
           )
         }
       </ComposerPrimitive.Unstable_TriggerPopoverItems>
-    </ComposerPrimitive.Unstable_TriggerPopover>
+    </MentionList>
   )
 }
diff --git a/surfsense_local/frontend/src/features/startup/startup-composer.tsx b/surfsense_local/frontend/src/features/startup/startup-composer.tsx
new file mode 100644
index 0000000000..ef546d2d93
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-composer.tsx
@@ -0,0 +1,108 @@
+import { useEffect, useState } from "react"
+
+import { Skeleton } from "@/components/ui/skeleton"
+import {
+  ComposerBox,
+  ComposerColumn,
+  ComposerControls,
+  ComposerText,
+} from "@/features/chat/composer-surface"
+import {
+  messageLabel,
+  startPlaceholder,
+} from "@/features/chat/composer-wording"
+import { QUESTION_MAX_CHARS } from "@/features/chat/question-limit"
+import {
+  MentionItem,
+  MentionItemDescription,
+  MentionList,
+} from "@/features/plugins/in-thread/mention-list"
+import {
+  MENTION_TRIGGER,
+  mentionAt,
+} from "@/features/plugins/in-thread/mention-matcher"
+import { cn } from "@/lib/utils"
+
+import { useStartupInput } from "./startup-input-context"
+
+// A name and a line about it per row, so the rows read as a list, not a block.
+const SKELETON_ROWS = [
+  ["w-24", "w-40"],
+  ["w-16", "w-32"],
+  ["w-28", "w-36"],
+] as const
+
+/** The `@` list while the plugins load, drawn as the real one is. */
+function MentionListSkeleton() {
+  return (
+    // Busy, with nothing to pick yet: a group, not a listbox of no options.
+    <MentionList role="group" aria-busy="true">
+      {SKELETON_ROWS.map(([name, about]) => (
+        <MentionItem key={name} aria-hidden="true">
+          {/* One line of the row's own text tall, whatever its size. */}
+          <span className="flex h-lh items-center">
+            <Skeleton className={cn("h-[0.7lh]", name)} />
+          </span>
+          <MentionItemDescription className="flex h-lh items-center">
+            <Skeleton className={cn("h-[0.7lh]", about)} />
+          </MentionItemDescription>
+        </MentionItem>
+      ))}
+    </MentionList>
+  )
+}
+
+/**
+ * A new chat's box while the app loads: it takes typing, keeps it for the
+ * composer that replaces it, and sends nothing, having nowhere to send yet.
+ * `mentions` is whether `@` names plugins in the chat it will become.
+ */
+export function StartupComposer({ mentions }: { mentions: boolean }) {
+  const startup = useStartupInput()
+  const [text, setText] = useState("")
+  const [caret, setCaret] = useState(0)
+
+  // A box gone mid-composition must not hold startup's end.
+  useEffect(() => () => startup?.setComposing(false), [startup])
+
+  const track = (box: HTMLTextAreaElement) => {
+    setText(box.value)
+    setCaret(box.selectionStart)
+    startup?.keep(box.value, box.selectionStart)
+  }
+
+  return (
+    <ComposerColumn data-composer-outline="center">
+      <ComposerBox placement="center">
+        {mentions && mentionAt(text, MENTION_TRIGGER, caret) ? (
+          <MentionListSkeleton />
+        ) : null}
+        <ComposerText
+          placement="center"
+          // As the composer it stands for does.
+          autoFocus
+          aria-label={messageLabel()}
+          placeholder={startPlaceholder()}
+          maxLength={QUESTION_MAX_CHARS}
+          onChange={(event) => track(event.currentTarget)}
+          onSelect={(event) => track(event.currentTarget)}
+          onKeyDown={(event) => {
+            if (
+              event.key === "Enter" &&
+              !event.shiftKey &&
+              !event.nativeEvent.isComposing
+            ) {
+              event.preventDefault()
+            }
+          }}
+          onCompositionStart={() => startup?.setComposing(true)}
+          onCompositionEnd={(event) => {
+            startup?.setComposing(false)
+            track(event.currentTarget)
+          }}
+        />
+        <ComposerControls />
+      </ComposerBox>
+    </ComposerColumn>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-draft-handover.tsx b/surfsense_local/frontend/src/features/startup/startup-draft-handover.tsx
new file mode 100644
index 0000000000..3deeaba4e5
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-draft-handover.tsx
@@ -0,0 +1,53 @@
+import {
+  unstable_useTriggerPopoverRootContextOptional,
+  useAui,
+  useAuiState,
+} from "@assistant-ui/react"
+import { useEffect, useRef, type RefObject } from "react"
+
+import { MENTION_TRIGGER } from "@/features/plugins/in-thread/mention-matcher"
+
+import type { StartupDraft } from "./startup-input"
+import { useStartupInput } from "./startup-input-context"
+
+/**
+ * Puts what was typed into the startup frame's box into the composer, caret
+ * and all. Rendered inside the composer's trigger root, whose `@` trigger
+ * keeps its own cursor: it is told where the caret is, so a mention the
+ * caret was left in opens the plugin list as typing it would have.
+ */
+export function StartupDraftHandover({
+  draft,
+  inputRef,
+}: {
+  draft: StartupDraft
+  inputRef: RefObject<HTMLTextAreaElement | null>
+}) {
+  const aui = useAui()
+  const startup = useStartupInput()
+  const triggers = unstable_useTriggerPopoverRootContextOptional()
+  const placed = useRef(false)
+  const shown = useAuiState(({ composer }) => composer.text === draft.text)
+
+  // After the composer's reset, an earlier sibling's effect, which empties it.
+  useEffect(() => {
+    startup?.take()
+    aui.thread.composer().setText(draft.text)
+  }, [aui, draft, startup])
+
+  // Once the text shows: a caret placed before then would land in nothing.
+  useEffect(() => {
+    const box = inputRef.current
+    if (!shown || placed.current || !box) return
+    placed.current = true
+    box.focus({ preventScroll: true })
+    box.setSelectionRange(draft.caret, draft.caret)
+    // Absent in Basic chats, which have no `@` list.
+    triggers
+      ?.getTriggers()
+      .get(MENTION_TRIGGER)
+      ?.resource.setCursorPosition(draft.caret)
+  }, [draft, inputRef, shown, triggers])
+
+  return null
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-frame.tsx b/surfsense_local/frontend/src/features/startup/startup-frame.tsx
index fe8956c145..61fd26385e 100644
--- a/surfsense_local/frontend/src/features/startup/startup-frame.tsx
+++ b/surfsense_local/frontend/src/features/startup/startup-frame.tsx
@@ -1,14 +1,9 @@
 import { useState } from "react"
 
 import {
-  THREAD_BODY_CLASS,
-  VIEWPORT_BOX_CLASS,
-  VIEWPORT_CLASS,
-  VIEWPORT_FOOTER_CLASS,
-  VIEWPORT_FOOTER_INNER_CLASS,
-  VIEWPORT_STYLE,
-} from "@/features/chat/chat-viewport-layout"
-import { ComposerOutline } from "@/features/chat/composer-outline"
+  ConversationBody,
+  ConversationScroll,
+} from "@/features/chat/conversation-scroll"
 import { ConversationColumn } from "@/features/chat/conversation-column"
 import { MessagesSkeleton } from "@/features/chat/messages-skeleton"
 import { ThreadWelcome } from "@/features/chat/thread-welcome"
@@ -21,12 +16,14 @@ import {
 import { RightPanel } from "@/features/dashboard/right-panel"
 import { SidebarFrame } from "@/features/dashboard/sidebar-frame"
 
+import { StartupComposer } from "./startup-composer"
 import { readStartupLayout } from "./startup-layout"
 
 /**
- * The dashboard's first frame without its data: the same columns, headings
- * and composer box, drawn by the dashboard's own parts, with nothing to press.
- * What needs data fades in over it once startup ends.
+ * The dashboard's first frame without its data: the same columns and
+ * headings, drawn by the dashboard's own parts. A new chat's box takes typing;
+ * an open thread shows its messages' skeleton and gets its composer with them.
+ * What needs data fades in once startup ends.
  */
 export function StartupFrame() {
   const [layout] = useState(readStartupLayout)
@@ -44,26 +41,18 @@ export function StartupFrame() {
         </SidebarColumn>
         <ChatColumn>
           <ConversationColumn>
-            <div className={THREAD_BODY_CLASS}>
-              <div className={VIEWPORT_BOX_CLASS}>
-                <div className={VIEWPORT_CLASS} style={VIEWPORT_STYLE}>
-                  {layout.composer === "center" ? (
-                    <ThreadWelcome
-                      composer={<ComposerOutline placement="center" />}
-                    />
-                  ) : (
-                    <>
-                      <MessagesSkeleton />
-                      <div className={VIEWPORT_FOOTER_CLASS}>
-                        <div className={VIEWPORT_FOOTER_INNER_CLASS}>
-                          <ComposerOutline placement="bottom" />
-                        </div>
-                      </div>
-                    </>
-                  )}
-                </div>
-              </div>
-            </div>
+            <ConversationBody>
+              <ConversationScroll>
+                {layout.composer === "center" ? (
+                  <ThreadWelcome
+                    composer={<StartupComposer mentions={layout.mentions} />}
+                  />
+                ) : (
+                  // The composer arrives with the thread it follows.
+                  <MessagesSkeleton />
+                )}
+              </ConversationScroll>
+            </ConversationBody>
           </ConversationColumn>
         </ChatColumn>
         <RightPanelRail open={layout.rightPanelOpen}>
diff --git a/surfsense_local/frontend/src/features/startup/startup-input-context.ts b/surfsense_local/frontend/src/features/startup/startup-input-context.ts
new file mode 100644
index 0000000000..465ee2b622
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-input-context.ts
@@ -0,0 +1,10 @@
+import { createContext, useContext } from "react"
+
+import type { StartupInput } from "./startup-input"
+
+// Null outside AppBootstrap, as in a component's own tests: nothing to hand over.
+export const StartupInputContext = createContext<StartupInput | null>(null)
+
+export function useStartupInput(): StartupInput | null {
+  return useContext(StartupInputContext)
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-input.test.ts b/surfsense_local/frontend/src/features/startup/startup-input.test.ts
new file mode 100644
index 0000000000..350b14a102
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-input.test.ts
@@ -0,0 +1,44 @@
+import { describe, expect, it } from "vitest"
+
+import { createStartupInput } from "./startup-input"
+
+describe("startup input", () => {
+  it("hands what was typed during startup to one composer only", () => {
+    const input = createStartupInput()
+    input.keep("find jobs with @ind", 19)
+    expect(input.peek()).toEqual({ text: "find jobs with @ind", caret: 19 })
+    input.take()
+    expect(input.peek()).toBeNull()
+  })
+
+  it("hands over nothing once the text was cleared", () => {
+    const input = createStartupInput()
+    input.keep("typo", 4)
+    input.keep("", 0)
+    expect(input.peek()).toBeNull()
+  })
+
+  it("is separate for each app, so nothing leaks between them", () => {
+    const one = createStartupInput()
+    one.keep("find jobs", 9)
+    expect(createStartupInput().peek()).toBeNull()
+  })
+
+  it("lets startup end at once while nothing is being composed", async () => {
+    await expect(createStartupInput().composingEnded()).resolves.toBeUndefined()
+  })
+
+  it("holds startup's end until a composition in progress ends", async () => {
+    const input = createStartupInput()
+    input.setComposing(true)
+    let ended = false
+    const waiting = input.composingEnded().then(() => {
+      ended = true
+    })
+    await Promise.resolve()
+    expect(ended).toBe(false)
+    input.setComposing(false)
+    await waiting
+    expect(ended).toBe(true)
+  })
+})
diff --git a/surfsense_local/frontend/src/features/startup/startup-input.ts b/surfsense_local/frontend/src/features/startup/startup-input.ts
new file mode 100644
index 0000000000..24425519ba
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/startup-input.ts
@@ -0,0 +1,43 @@
+export type StartupDraft = { text: string; caret: number }
+
+/**
+ * What was typed into the startup frame's box, held for the composer that
+ * replaces it, and whether an input method is mid-composition there. One per
+ * app, made by AppBootstrap: the frame and the composer are siblings in its
+ * tree, never the same component.
+ */
+export type StartupInput = {
+  keep: (text: string, caret: number) => void
+  peek: () => StartupDraft | null
+  // After the composer has the text, so a composer mounted later starts empty.
+  take: () => void
+  setComposing: (active: boolean) => void
+  // Handing over mid-composition would cut the input method's unconverted text.
+  composingEnded: () => Promise<void>
+}
+
+export function createStartupInput(): StartupInput {
+  let draft: StartupDraft | null = null
+  let composing = false
+  let waiting: (() => void)[] = []
+  return {
+    keep: (text, caret) => {
+      draft = text ? { text, caret } : null
+    },
+    peek: () => draft,
+    take: () => {
+      draft = null
+    },
+    setComposing: (active) => {
+      composing = active
+      if (active) return
+      const resolved = waiting
+      waiting = []
+      for (const resolve of resolved) resolve()
+    },
+    composingEnded: () =>
+      composing
+        ? new Promise<void>((resolve) => waiting.push(resolve))
+        : Promise.resolve(),
+  }
+}
diff --git a/surfsense_local/frontend/src/features/startup/startup-layout.test.ts b/surfsense_local/frontend/src/features/startup/startup-layout.test.ts
index 70d4baeb9d..9b4735e411 100644
--- a/surfsense_local/frontend/src/features/startup/startup-layout.test.ts
+++ b/surfsense_local/frontend/src/features/startup/startup-layout.test.ts
@@ -9,6 +9,7 @@ describe("startup layout", () => {
     expect(readStartupLayout()).toEqual({
       composer: "center",
       rightPanelOpen: true,
+      mentions: false,
     })
   })
 
diff --git a/surfsense_local/frontend/src/features/startup/startup-layout.ts b/surfsense_local/frontend/src/features/startup/startup-layout.ts
index 5a1837c909..2770a26c25 100644
--- a/surfsense_local/frontend/src/features/startup/startup-layout.ts
+++ b/surfsense_local/frontend/src/features/startup/startup-layout.ts
@@ -1,10 +1,13 @@
 import { readLastThread } from "@/features/chat/last-thread"
+import { readNextNewChatMode } from "@/features/chat/modes/next-new-chat-mode-store"
 import { readRightPanelOpen } from "@/features/dashboard/chrome-prefs"
 import { readStoredWorkspaceId } from "@/features/workspaces/last-workspace"
 
 export type StartupLayout = {
   composer: "center" | "bottom"
   rightPanelOpen: boolean
+  // Whether `@` in a new chat's box names plugins: only in Agentic chats.
+  mentions: boolean
 }
 
 /**
@@ -20,5 +23,6 @@ export function readStartupLayout(): StartupLayout {
         ? "bottom"
         : "center",
     rightPanelOpen: readRightPanelOpen(),
+    mentions: readNextNewChatMode() === "agentic",
   }
 }
diff --git a/surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx b/surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx
index 1062484929..3aa9f929bf 100644
--- a/surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx
+++ b/surfsense_local/frontend/src/features/startup/startup-reveal.test.tsx
@@ -1,13 +1,18 @@
-import { act, cleanup, render, screen } from "@testing-library/react"
+import { cleanup, fireEvent, render, screen } from "@testing-library/react"
 import userEvent from "@testing-library/user-event"
 import { useState } from "react"
-import { afterEach, describe, expect, it, vi } from "vitest"
+import { afterEach, describe, expect, it } from "vitest"
 
 import { StartupRevealProvider } from "./startup-reveal-provider"
 import { useStartupReveal } from "./use-startup-reveal"
 
 function Region({ name }: { name: string }) {
-  return <div data-testid={name} className={useStartupReveal()} />
+  const reveal = useStartupReveal()
+  return (
+    <div data-testid={name} {...reveal}>
+      <span data-testid={`${name}-contents`} />
+    </div>
+  )
 }
 
 function LateRegion() {
@@ -21,10 +26,7 @@ function LateRegion() {
   )
 }
 
-afterEach(() => {
-  cleanup()
-  vi.useRealTimers()
-})
+afterEach(() => cleanup())
 
 describe("startup reveal", () => {
   it("fades in what the dashboard first draws after a visible wait", () => {
@@ -45,14 +47,13 @@ describe("startup reveal", () => {
     expect(screen.getByTestId("region").className).not.toContain("ss-reveal")
   })
 
-  it("stops fading once the fade is over, so later contents arrive as usual", () => {
-    vi.useFakeTimers()
+  it("stops fading once its contents have faded in, so later ones arrive as usual", () => {
     render(
       <StartupRevealProvider reveal>
         <Region name="region" />
       </StartupRevealProvider>
     )
-    act(() => vi.advanceTimersByTime(1000))
+    fireEvent.animationEnd(screen.getByTestId("region-contents"))
     expect(screen.getByTestId("region").className).not.toContain("ss-reveal")
   })
 
diff --git a/surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts b/surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts
index 39ec02afbb..1307a9b060 100644
--- a/surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts
+++ b/surfsense_local/frontend/src/features/startup/startup-screen-imports.test.ts
@@ -1,3 +1,4 @@
+import ts from "typescript"
 import { describe, expect, it } from "vitest"
 
 // The startup screen ships in the first chunk, before the dashboard's. If it
@@ -31,7 +32,46 @@ function resolveLocal(from: string, specifier: string): string | null {
   return null
 }
 
-// Every module the file loads at run time; type-only imports are erased.
+function isTypeOnly(node: ts.ImportDeclaration | ts.ExportDeclaration) {
+  if (ts.isExportDeclaration(node)) return node.isTypeOnly
+  const clause = node.importClause
+  if (!clause) return false
+  if (clause.isTypeOnly) return true
+  // `import { type A, type B }`, with nothing else, loads nothing either.
+  const bindings = clause.namedBindings
+  return (
+    !clause.name &&
+    bindings !== undefined &&
+    ts.isNamedImports(bindings) &&
+    bindings.elements.length > 0 &&
+    bindings.elements.every((element) => element.isTypeOnly)
+  )
+}
+
+/** The modules a file loads when it loads: static, value imports only. */
+function loadedBy(file: string): string[] {
+  const source = ts.createSourceFile(
+    file,
+    SOURCES[file] ?? "",
+    ts.ScriptTarget.Latest,
+    false,
+    file.endsWith(".tsx") ? ts.ScriptKind.TSX : ts.ScriptKind.TS
+  )
+  const specifiers: string[] = []
+  for (const statement of source.statements) {
+    if (
+      (ts.isImportDeclaration(statement) ||
+        ts.isExportDeclaration(statement)) &&
+      statement.moduleSpecifier &&
+      ts.isStringLiteral(statement.moduleSpecifier) &&
+      !isTypeOnly(statement)
+    ) {
+      specifiers.push(statement.moduleSpecifier.text)
+    }
+  }
+  return specifiers
+}
+
 function reachable(entry: string): string[] {
   const seen = new Set<string>()
   const queue = [entry]
@@ -40,10 +80,7 @@ function reachable(entry: string): string[] {
     const file = queue.pop()!
     if (seen.has(file)) continue
     seen.add(file)
-    const imports = (SOURCES[file] ?? "").matchAll(
-      /^\s*import\s+(?!type\b)(?:[^"';]*?\sfrom\s+)?["']([^"']+)["']/gm
-    )
-    for (const [, specifier] of imports) {
+    for (const specifier of loadedBy(file)) {
       if (specifier.startsWith(".") || specifier.startsWith("@/")) {
         const local = resolveLocal(file, specifier)
         if (local) {
@@ -65,6 +102,13 @@ describe("startup screen's imports", () => {
     expect(reached.some((module) => FORBIDDEN[2]!.test(module))).toBe(true)
   })
 
+  it("skip what only names types, which loads nothing", () => {
+    // It names assistant-ui's matcher type only.
+    expect(
+      loadedBy("/src/features/plugins/in-thread/mention-matcher.ts")
+    ).toEqual([])
+  })
+
   it("reach neither the chat runtime nor any feature's API client", () => {
     const reached = reachable("/src/features/startup/startup-screen.tsx")
     expect(reached.length).toBeGreaterThan(10)
diff --git a/surfsense_local/frontend/src/features/startup/startup-screen.tsx b/surfsense_local/frontend/src/features/startup/startup-screen.tsx
index 2a596cf719..48f05f6df7 100644
--- a/surfsense_local/frontend/src/features/startup/startup-screen.tsx
+++ b/surfsense_local/frontend/src/features/startup/startup-screen.tsx
@@ -9,16 +9,16 @@ import { StartupFrame } from "./startup-frame"
  * onboarding never opens over a glimpse of the dashboard.
  */
 export function StartupScreen({ frame }: { frame: boolean }) {
+  const label = intl.formatMessage({
+    id: "app_bootstrap_loader_aria",
+    defaultMessage: "Starting SurfSense",
+  })
   return (
-    <DashboardShell
-      role="status"
-      aria-busy="true"
-      aria-label={intl.formatMessage({
-        id: "app_bootstrap_loader_aria",
-        defaultMessage: "Starting SurfSense",
-      })}
-      className="select-none"
-    >
+    // Busy, not a status region: the frame holds a box to type in.
+    <DashboardShell aria-busy="true" aria-label={label}>
+      <p role="status" className="sr-only">
+        {label}
+      </p>
       {frame ? <StartupFrame /> : null}
     </DashboardShell>
   )
diff --git a/surfsense_local/frontend/src/features/startup/use-startup-draft.ts b/surfsense_local/frontend/src/features/startup/use-startup-draft.ts
new file mode 100644
index 0000000000..031718e0d7
--- /dev/null
+++ b/surfsense_local/frontend/src/features/startup/use-startup-draft.ts
@@ -0,0 +1,16 @@
+import { useState } from "react"
+
+import type { StartupDraft } from "./startup-input"
+import { useStartupInput } from "./startup-input-context"
+
+/**
+ * What was typed into the startup frame's box, if anything, read once when
+ * the composer mounts: a draft that arrives later belongs to no composer on
+ * screen. The composer leaves its own mount focus off when there is one, as
+ * that focus moves the caret to the end.
+ */
+export function useStartupDraft(): StartupDraft | null {
+  const startup = useStartupInput()
+  const [draft] = useState(() => startup?.peek() ?? null)
+  return draft
+}
diff --git a/surfsense_local/frontend/src/features/startup/use-startup-reveal.ts b/surfsense_local/frontend/src/features/startup/use-startup-reveal.ts
index f275b20dd7..e646da59f8 100644
--- a/surfsense_local/frontend/src/features/startup/use-startup-reveal.ts
+++ b/surfsense_local/frontend/src/features/startup/use-startup-reveal.ts
@@ -1,21 +1,31 @@
-import { useContext, useEffect, useState } from "react"
+import { useCallback, useContext, useState } from "react"
 
 import { StartupRevealContext } from "./startup-reveal-context"
 
-/** `.ss-reveal` in `index.css` runs 200 ms; the class is dropped after it ends. */
-const REVEAL_CLEAR_MS = 400
+export type StartupReveal = {
+  className?: string
+  ref?: (element: HTMLElement | null) => (() => void) | undefined
+}
 
 /**
- * The class for a container whose contents fade in at startup while the
- * container stays put: the loading frame drew the container already.
+ * Props for a container whose contents fade in at startup while it stays
+ * put: the loading frame drew it already. Spread them, and merge `className`.
+ * The fade ends when its contents' animation does, so contents added later
+ * arrive as usual; under reduced motion nothing animates and nothing shows.
  */
-export function useStartupReveal(): string {
+export function useStartupReveal(): StartupReveal {
   const reveal = useContext(StartupRevealContext)
   const [fading, setFading] = useState(reveal)
-  useEffect(() => {
-    if (!fading) return
-    const timer = setTimeout(() => setFading(false), REVEAL_CLEAR_MS)
-    return () => clearTimeout(timer)
-  }, [fading])
-  return fading ? "ss-reveal" : ""
+  // A native listener: React names animation events by what the page
+  // supports, so its own handler is not one name everywhere.
+  const ref = useCallback((element: HTMLElement | null) => {
+    if (!element) return undefined
+    // The contents animate, not the container; their end bubbles up here.
+    const ended = (event: Event) => {
+      if (event.target !== element) setFading(false)
+    }
+    element.addEventListener("animationend", ended)
+    return () => element.removeEventListener("animationend", ended)
+  }, [])
+  return fading ? { className: "ss-reveal", ref } : {}
 }
diff --git a/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx b/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx
index 2f357d2fd0..ac1f6b2eae 100644
--- a/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx
+++ b/surfsense_local/frontend/src/features/workspaces/workspace-rail.tsx
@@ -182,9 +182,10 @@ export function WorkspaceRail({
 
   return (
     <nav
+      {...reveal}
       className={cn(
         "flex h-full flex-col items-center bg-app-shell py-3 text-sidebar-foreground",
-        reveal
+        reveal.className
       )}
       aria-label={intl.formatMessage({
         id: "workspaces_rail_aria",
diff --git a/surfsense_local/frontend/src/index.css b/surfsense_local/frontend/src/index.css
index d58f99b9b4..8979a78735 100644
--- a/surfsense_local/frontend/src/index.css
+++ b/surfsense_local/frontend/src/index.css
@@ -423,12 +423,13 @@
   }
 }
 
-.ss-reveal > * {
+/* `data-reveal-steady` stays solid: the composer's text, typed during startup. */
+.ss-reveal > :not([data-reveal-steady]) {
   animation: ss-reveal-in 200ms cubic-bezier(0.215, 0.61, 0.355, 1);
 }
 
 @media (prefers-reduced-motion: reduce) {
-  .ss-reveal > * {
+  .ss-reveal > :not([data-reveal-steady]) {
     animation: none;
   }
 }

From a2e30f613963bdbe8d232bf2828edd7e65acb1bc Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 01:34:25 +0530
Subject: [PATCH 03/12] fix(local): run the sidebar and artifact lists to the
 bottom edge

---
 .../frontend/src/features/dashboard/left-sidebar.tsx            | 2 +-
 surfsense_local/frontend/src/features/dashboard/right-panel.tsx | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx b/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
index cbc0db4832..4093780110 100644
--- a/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/left-sidebar.tsx
@@ -206,7 +206,7 @@ export const LeftSidebar = memo(function LeftSidebar({
       <div
         {...reveal}
         className={cn(
-          "flex min-h-0 min-w-0 flex-1 flex-col overflow-hidden px-3 pt-5 pb-2",
+          "flex min-h-0 min-w-0 flex-1 flex-col overflow-hidden px-3 pt-5",
           reveal.className
         )}
       >
diff --git a/surfsense_local/frontend/src/features/dashboard/right-panel.tsx b/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
index f810a3cb52..48a6049746 100644
--- a/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/right-panel.tsx
@@ -50,7 +50,7 @@ export function RightPanel({
       <div
         {...reveal}
         className={cn(
-          "min-h-0 min-w-0 flex-1 overflow-hidden px-3 pb-2",
+          "min-h-0 min-w-0 flex-1 overflow-hidden px-3",
           reveal.className
         )}
       >

From d60d0a4f7f46602a93112c994b99fbfa691ca508 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 01:52:07 +0530
Subject: [PATCH 04/12] fix(local): centre the title-bar controls in the strip
 on every OS

---
 .../frontend/src/features/dashboard/dashboard-columns.tsx  | 2 +-
 surfsense_local/frontend/src/index.css                     | 7 ++++++-
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx b/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx
index 90de0ee814..846bcb200f 100644
--- a/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/dashboard-columns.tsx
@@ -43,7 +43,7 @@ export function DashboardColumns({
   return (
     <section
       ref={sectionRef}
-      className="my-2 mr-2 flex min-h-0 min-w-0 overflow-hidden rounded-[16px] border bg-background shadow-sm"
+      className="my-(--columns-gap) mr-2 flex min-h-0 min-w-0 overflow-hidden rounded-[16px] border bg-background shadow-sm"
     >
       <ColumnsContext.Provider value={columns}>
         {children}
diff --git a/surfsense_local/frontend/src/index.css b/surfsense_local/frontend/src/index.css
index 8979a78735..55022e73e8 100644
--- a/surfsense_local/frontend/src/index.css
+++ b/surfsense_local/frontend/src/index.css
@@ -65,6 +65,9 @@
 :root {
   color-scheme: light;
   --titlebar-height: 0px;
+  /* Between the title bar and the dashboard's columns; the title-bar
+     controls centre across it too. */
+  --columns-gap: 0.5rem;
   --app-shell: #f3f2ee;
   --background: #f7f6f2;
   --foreground: #1e1e1e;
@@ -255,7 +258,9 @@
   align-items: center;
   gap: 0.375rem;
   margin-inline-start: auto;
-  transform: translateY(2px);
+  /* Centred in the strip the eye sees, title bar plus the gap above the
+     columns, on every OS: the bar centres them, this adds half the gap. */
+  transform: translateY(calc(var(--columns-gap) / 2));
   app-region: no-drag;
   -webkit-app-region: no-drag;
 }

From ed16415d6ee35d55a53add72a7a88e3515e94140 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 12:27:30 +0530
Subject: [PATCH 05/12] fix(local): save a chat mode when it is picked

A pick was only remembered once a chat opened, so choosing Basic (Q&A) and
refreshing went back to the model's default. It is now saved per model at the
pick, and a refused Agentic is dropped when a local model's facts are read.
---
 docs/architecture/chat.md                     |  4 +-
 docs/architecture/model-capabilities.md       |  7 +-
 .../backend/modules/agent/engine_choice.py    | 28 +++----
 .../backend/modules/chat/router.py            |  3 -
 .../backend/modules/chat/schemas.py           |  6 +-
 .../modules/llm/capability/modes/__init__.py  |  2 +
 .../llm/capability/modes/remembered.py        | 32 ++++++--
 .../modules/llm/capability/modes/router.py    | 49 ++++++++++++
 .../backend/modules/llm/capability/read.py    | 22 +++---
 surfsense_local/backend/modules/llm/router.py |  2 +
 .../integration/agent/test_agent_threads.py   | 14 ++--
 .../integration/agent/test_engine_choice.py   | 38 ++++-----
 .../integration/agent/test_thread_modes.py    | 59 ++------------
 .../integration/llm/test_capability_routes.py |  8 +-
 .../integration/llm/test_pick_chat_mode.py    | 79 +++++++++++++++++++
 .../llm/capability/test_new_chat_modes.py     | 14 ----
 .../frontend/src/features/chat/api.ts         |  2 -
 .../features/chat/modes/mode-picker.test.tsx  | 29 +++++++
 .../src/features/chat/modes/mode-picker.tsx   | 13 ++-
 .../src/features/chat/modes/new-chat-mode.ts  | 17 ++--
 .../src/features/chat/modes/save-mode-pick.ts | 52 ++++++++++++
 .../features/chat/use-chat-runtime.test.tsx   | 20 +----
 .../dashboard/dashboard-page.test.tsx         |  6 +-
 surfsense_local/frontend/translations/de.json |  1 +
 surfsense_local/frontend/translations/en.json |  1 +
 surfsense_local/frontend/translations/es.json |  1 +
 surfsense_local/frontend/translations/fr.json |  1 +
 surfsense_local/frontend/translations/hi.json |  1 +
 surfsense_local/frontend/translations/ja.json |  1 +
 surfsense_local/frontend/translations/ko.json |  1 +
 .../frontend/translations/pt-BR.json          |  1 +
 surfsense_local/frontend/translations/ru.json |  1 +
 .../frontend/translations/zh-CN.json          |  1 +
 33 files changed, 327 insertions(+), 189 deletions(-)
 create mode 100644 surfsense_local/backend/modules/llm/capability/modes/router.py
 create mode 100644 surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py
 create mode 100644 surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts

diff --git a/docs/architecture/chat.md b/docs/architecture/chat.md
index 4a22bdc0f4..1731b62e79 100644
--- a/docs/architecture/chat.md
+++ b/docs/architecture/chat.md
@@ -9,7 +9,7 @@ A chat thread belongs to a workspace. Each user message retrieves its own contex
 
 | Method | Path | Does |
 |---|---|---|
-| `POST` | `/workspaces/{workspace_id}/chat/threads` | open a thread, with an optional `source_scope`, `mode`, `basic` or `agentic`, and `remember`, which makes the mode the model's default ([Agent threads](#agent-threads)); the title defaults to "New chat" |
+| `POST` | `/workspaces/{workspace_id}/chat/threads` | open a thread, with an optional `source_scope` and `mode`, `basic` or `agentic` ([Agent threads](#agent-threads)); the title defaults to "New chat" |
 | `GET` | `/workspaces/{workspace_id}/chat/threads` | list, newest first, each with `running` while a reply is being generated for it, and `run_state`, `{state, position}`: `queued` with its place in line, `running`, or `needs-approval` while an agent turn waits on the user; `null` when nothing runs |
 | `PATCH` | `/chat/threads/{thread_id}` | rename, 1 to 200 characters |
 | `DELETE` | `/chat/threads/{thread_id}` | delete; a running reply is stopped first, waiting up to 5 seconds, then its messages cascade, its artifacts stay with the thread cleared, and its images' folder goes after the commit |
@@ -267,7 +267,7 @@ Interface strings are `dictation_<surface>_<purpose>`, and a server code is `dic
 
 ## Agent threads
 
-How the agent runs is in [agent](agent.md). A thread opened with `mode: "agentic"` is the agent's, and one opened with `mode: "basic"` the chat's; with no `mode` it takes the selected model's default ([model capabilities](model-capabilities.md#the-modes), [agent](agent.md#which-threads-get-it)). Agentic asked for and kept out by a technical gate is refused with `409`, `detail.code` `no_model`, `agent_not_installed`, `tool_calls_unsupported` or `window_below_floor`, and no thread is made. Opening an agent thread starts opencode if it is not running, waits until it serves SurfSense's configuration, and opens an opencode session under the thread's title; the session's id and the engine, `agent_engine` `opencode`, are stored on the thread, and `ThreadRead` reports `uses_agent`, true when the thread has an engine. When opencode does not become ready, a thread that asked for Agentic is removed and the request answers `503`, `detail.code` `agent_unavailable`; one that asked for nothing opens as a chat. A mode sent with `remember: true`, which the composer sends only for a mode the user picked, is remembered as the model's default. Which engine runs a thread's turns, and how it moves when the selected model changes, is in [agent](agent.md#two-engines).
+How the agent runs is in [agent](agent.md). A thread opened with `mode: "agentic"` is the agent's, and one opened with `mode: "basic"` the chat's; with no `mode` it takes the selected model's default ([model capabilities](model-capabilities.md#the-modes), [agent](agent.md#which-threads-get-it)). Agentic asked for and kept out by a technical gate is refused with `409`, `detail.code` `no_model`, `agent_not_installed`, `tool_calls_unsupported` or `window_below_floor`, and no thread is made. Opening an agent thread starts opencode if it is not running, waits until it serves SurfSense's configuration, and opens an opencode session under the thread's title; the session's id and the engine, `agent_engine` `opencode`, are stored on the thread, and `ThreadRead` reports `uses_agent`, true when the thread has an engine. When opencode does not become ready, a thread that asked for Agentic is removed and the request answers `503`, `detail.code` `agent_unavailable`; one that asked for nothing opens as a chat. Opening a thread never changes the model's default; the composer's switch saves a pick itself ([model capabilities](model-capabilities.md#the-modes)). Which engine runs a thread's turns, and how it moves when the selected model changes, is in [agent](agent.md#two-engines).
 
 The same routes then reach the agent ([`modules/agent/agent_threads/`](../../surfsense_local/backend/modules/agent/agent_threads/)):
 
diff --git a/docs/architecture/model-capabilities.md b/docs/architecture/model-capabilities.md
index 5b819c7e2b..01d78a7290 100644
--- a/docs/architecture/model-capabilities.md
+++ b/docs/architecture/model-capabilities.md
@@ -50,7 +50,7 @@ A chat is Basic (Q&A), the chat engine's retrieval answers with Studio beside it
 - **`agentic_allowed`**, and when not, **`blocked`**: a technical gate only, never a score ([ADR 0044](../adr/0044-a-score-sets-the-default-mode-never-whether-agentic-is-allowed.md)). `agent_not_installed` when no opencode is staged, except on a Claude plan, whose agent is the user's Claude Code; `tool_calls_unsupported` when the remote catalog states `tool_call: false`, or llama-server does not report `supports_tool_calls` for a local model's template; `window_below_floor` when the stated window is under 32,768 tokens ([`agent_gate.py`](../../surfsense_local/backend/modules/llm/capability/agent_gate.py)). A catalog that says nothing, and a runtime that cannot be read, allow it. A local model's tool calls and window are read when a chat starts, since reading them would load the model, so the selection read allows it until then; a chat refused by one of these gates disables Agentic for that model in the composer until the app restarts.
 - **`default_mode`**: the mode the user last chose for a new chat on this model; else Agentic for `agent` and `agent_limited`, which hold only on a remote host; else Basic. A blocked Agentic defaults to Basic.
 - **`reason`**, what the switch says beside Agentic: `measured_pass` or `measured_near` with `passed` and `counted`; `measured_below`, a low scorer, with the same; `assumed`; `local_copy`, a copy on the user's own machine or network of a model whose row passed on a remote host, with that `host`; `untested`. An assumed row says `assumed` on a local copy too, since it was run nowhere, and the capability's own `reason` is `assumed` there rather than `measured_elsewhere`.
-- **`remembered_mode`**, under `chat_modes` in the text selection's `settings`, one entry per provider, connection and model name ([`modes/remembered.py`](../../surfsense_local/backend/modules/llm/capability/modes/remembered.py)), written when a chat is opened with a mode the user chose (`remember: true`). A default sent back unchosen is not written, so a later score or the developer switch turned off still decides. The entries stay when the slot takes another model, so a model's choice comes back with it. The `agent_trial` the modes replaced reads as its model's Agentic, kept as such when the slot changes, and the first write drops it.
+- **`remembered_mode`**, under `chat_modes` in the text selection's `settings`, one entry per provider, connection and model name ([`modes/remembered.py`](../../surfsense_local/backend/modules/llm/capability/modes/remembered.py)), written when the user picks a mode in the composer's switch (`PUT /llm/selection/text_gen/mode`). Opening a chat writes nothing, so a default the user never touched follows a later score, or the developer switch turned off. A saved Agentic that a local model's runtime then refuses by a gate is dropped, so the model's own default holds again. The entries stay when the slot takes another model, so a model's choice comes back with it. The `agent_trial` the modes replaced reads as its model's Agentic, kept as such when the slot changes, and the first write drops it.
 
 `SURFSENSE_LOCAL_AGENT_UNTESTED_MODELS=1` makes Agentic the default for every model the user has not picked a mode for and lifts the window floor; a stated no on tool calls, or no opencode, still keeps Agentic out.
 
@@ -58,11 +58,12 @@ A chat is Basic (Q&A), the chat engine's retrieval answers with Studio beside it
 
 - `SelectionRead.capability` on `GET` and `PUT /llm/selection/text_gen` (null for other slots): `level`, `label_key` (the ICU select branch), `reason` (`code`, `values`), `note`, `measured` (`key`, `suite`, `assumed`, `suite_version`, `measured_on`, `provider`, `host`, `reads_images`, `passed`, `counted`, `provisional`, or null) and `modes` (`agentic_allowed`, `blocked`, `default_mode`: `basic` or `agentic`, `reason` (`code`, `values`), `remembered_mode`), and `studio`: `full` or `simple`, the Studio path below.
 - `capability_level` on each row of `GET /llm/connections/{id}/models` and `GET /llm/providers/llamacpp/models` that can fill the chat slot.
-- `POST /workspaces/{id}/chat/threads` takes `mode` and `remember` ([chat](chat.md#agent-threads)).
+- `PUT /llm/selection/text_gen/mode` with `{"mode": "basic" | "agentic"}` saves the pick as the selected model's `remembered_mode` and returns its `modes`; `404` with no chat model chosen, `409` with the gate's `detail.code` for Agentic on a model a gate keeps out ([`modes/router.py`](../../surfsense_local/backend/modules/llm/capability/modes/router.py)).
+- `POST /workspaces/{id}/chat/threads` takes `mode` ([chat](chat.md#agent-threads)).
 
 In dev builds only, the model picker tags each chat model by how it did in the Agentic tests: "Agentic", "Agentic, may need nudges", "Low Agentic score" or "Not tested"; production shows no tag. In every build, Settings › Models shows **What this model can do** under the chat model in use ([`model-capability.tsx`](../../surfsense_local/frontend/src/features/models/capability/model-capability.tsx)), closed until opened, so a rating never reads as a verdict beside the model's name. Opened, it says in plain words what Chat offers (both modes and how well, or Basic only and why), what Studio makes (`studio`: full designed pages and Office files, or simple ones and what a stronger model adds), and, last and small, where the rating came from: the cases passed and the day, "From a short test" while the list is provisional, or why it is untested. The row's English note is not shown.
 
-The composer's switch ([`mode-picker.tsx`](../../surfsense_local/frontend/src/features/chat/modes/mode-picker.tsx)) is a menu of "Basic (Q&A)" and "Agentic", each with one line on what it does. On a new chat it starts at the model's `default_mode` and follows the model when it changes; a pick holds for that model for the session, in a context above the open workspace. The new thread's `mode` is always sent, with `remember: true` only when the user picked it, so only a pick becomes the model's default. An Agentic that did not pass carries a tag ("Not tested", "Low score", "May need nudges", "Local copy", "Expected to pass"); the `reason`, or a gate, shows in a tooltip beside the Agentic row on hover or focus and is read with the row, in both menus: "Passed 2 of 8 Agentic tests, so it may stop early or make mistakes.", "Not tested with this model yet, so it may stop early or make mistakes.", "Passed on its full-size version. A local copy may do worse.", or that a flagship is assumed. Nothing is said under the composer. A gated Agentic is disabled with its reason, such as "This model can’t use tools." In an open chat, under "This chat’s mode", the switch lists the same two rows with the chat's mode checked; the other is marked "New chat", and picking it picks that mode and opens a new chat. A refused Agentic chat is worded by its code in a toast, and the question and images go back into the composer. A refusal by a gate the selection could not say, such as a local model's tool calls, disables Agentic with its reason for that model for the session ([`new-chat-mode.ts`](../../surfsense_local/frontend/src/features/chat/modes/new-chat-mode.ts)), so the next chat starts in Basic (Q&A) rather than asking again. The agent reads no attached image yet, so an Agentic chat, new or open, takes none ([chat](chat.md)).
+The composer's switch ([`mode-picker.tsx`](../../surfsense_local/frontend/src/features/chat/modes/mode-picker.tsx)) is a menu of "Basic (Q&A)" and "Agentic", each with one line on what it does. On a new chat it starts at the model's `default_mode` and follows the model when it changes; a pick shows at once, from a context above the open workspace, and is saved as that model's default, so a refresh or a restart keeps it ([`save-mode-pick.ts`](../../surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts)). When the save fails, the pick holds for the session and a toast says so. The new thread's `mode` is always sent. An Agentic that did not pass carries a tag ("Not tested", "Low score", "May need nudges", "Local copy", "Expected to pass"); the `reason`, or a gate, shows in a tooltip beside the Agentic row on hover or focus and is read with the row, in both menus: "Passed 2 of 8 Agentic tests, so it may stop early or make mistakes.", "Not tested with this model yet, so it may stop early or make mistakes.", "Passed on its full-size version. A local copy may do worse.", or that a flagship is assumed. Nothing is said under the composer. A gated Agentic is disabled with its reason, such as "This model can’t use tools." In an open chat, under "This chat’s mode", the switch lists the same two rows with the chat's mode checked; the other is marked "New chat", and picking it picks that mode and opens a new chat. A refused Agentic chat is worded by its code in a toast, and the question and images go back into the composer. A refusal by a gate the selection could not say, such as a local model's tool calls, disables Agentic with its reason for that model for the session ([`new-chat-mode.ts`](../../surfsense_local/frontend/src/features/chat/modes/new-chat-mode.ts)), so the next chat starts in Basic (Q&A) rather than asking again. The agent reads no attached image yet, so an Agentic chat, new or open, takes none ([chat](chat.md)).
 
 ## Known gaps
 
diff --git a/surfsense_local/backend/modules/agent/engine_choice.py b/surfsense_local/backend/modules/agent/engine_choice.py
index 1f35e41272..77affe5cde 100644
--- a/surfsense_local/backend/modules/agent/engine_choice.py
+++ b/surfsense_local/backend/modules/agent/engine_choice.py
@@ -18,9 +18,9 @@
 from modules.llm.capability.modes import (
     ChatMode,
     NewChatModes,
+    forget_agentic,
     mode_entry,
     new_chat_modes,
-    remember_mode,
 )
 from modules.llm.model_type import ModelType
 from modules.llm.models import SelectedModel
@@ -31,7 +31,6 @@
     "AgenticRefusedError",
     "NewThreadMode",
     "new_thread_mode",
-    "remember_thread_mode",
     "selected_model_can_run_agent",
 ]
 
@@ -70,9 +69,6 @@ class _Chosen:
 @dataclass(frozen=True)
 class NewThreadMode:
     mode: ChatMode
-    # The model it was decided on, as its remembered mode is kept; None with no model.
-    # Read with the choice: the slot may hold another once the agent has started.
-    entry: str | None
 
 
 async def new_thread_mode(
@@ -87,19 +83,21 @@ async def new_thread_mode(
     if chosen is None:
         if requested is ChatMode.AGENTIC:
             raise AgenticRefusedError("no_model")
-        return NewThreadMode(ChatMode.BASIC, None)
+        return NewThreadMode(ChatMode.BASIC)
     mode = requested or chosen.modes.default_mode
     if mode is ChatMode.BASIC:
-        return NewThreadMode(mode, chosen.entry)
+        return NewThreadMode(mode)
     blocked = chosen.modes.blocked
     if blocked is None and chosen.local_runtime:
         # Read only now: it loads the model, which a Basic chat may not need yet.
         blocked = _gate(await local_facts(chosen.name))
+        if blocked is not None:
+            await transact(session, _forget_agentic, chosen.entry)
     if blocked is None:
-        return NewThreadMode(ChatMode.AGENTIC, chosen.entry)
+        return NewThreadMode(ChatMode.AGENTIC)
     if requested is ChatMode.AGENTIC:
         raise AgenticRefusedError(blocked)
-    return NewThreadMode(ChatMode.BASIC, chosen.entry)
+    return NewThreadMode(ChatMode.BASIC)
 
 
 async def selected_model_can_run_agent(session: Session) -> bool:
@@ -118,14 +116,6 @@ async def selected_model_can_run_agent(session: Session) -> bool:
     return _gate(facts) is None
 
 
-async def remember_thread_mode(session: Session, entry: str, mode: ChatMode) -> None:
-    """The mode the user chose for a new chat becomes that model's default for new ones.
-
-    `entry` is `NewThreadMode.entry`, the model the chat's mode was decided on.
-    """
-    await transact(session, _remember, entry, mode)
-
-
 def _gate(facts: ToolFacts) -> str | None:
     return gate_block(
         facts, window_floor=not get_agent_settings().agent_untested_models
@@ -147,7 +137,7 @@ def _chosen(session: Session) -> _Chosen | None:
     )
 
 
-def _remember(session: Session, entry: str, mode: ChatMode) -> None:
+def _forget_agentic(session: Session, entry: str) -> None:
     selected = session.get(SelectedModel, ModelType.TEXT_GEN)
     if selected is not None:
-        remember_mode(selected, mode, entry)
+        forget_agentic(selected, entry)
diff --git a/surfsense_local/backend/modules/chat/router.py b/surfsense_local/backend/modules/chat/router.py
index ed249ad396..0e0b7ba3b7 100644
--- a/surfsense_local/backend/modules/chat/router.py
+++ b/surfsense_local/backend/modules/chat/router.py
@@ -21,7 +21,6 @@
 from modules.agent.engine_choice import (
     AgenticRefusedError,
     new_thread_mode,
-    remember_thread_mode,
 )
 from modules.agent.engines.both_sessions import thread_turns_of
 from modules.agent.engines.handover.take_over import take_over
@@ -132,8 +131,6 @@ async def create_thread(
                     "Basic (Q&A) chat.",
                 },
             )
-    if payload.mode is not None and payload.remember and opening.entry is not None:
-        await remember_thread_mode(session, opening.entry, payload.mode)
     return thread
 
 
diff --git a/surfsense_local/backend/modules/chat/schemas.py b/surfsense_local/backend/modules/chat/schemas.py
index ccdf9157dc..f84042eb02 100644
--- a/surfsense_local/backend/modules/chat/schemas.py
+++ b/surfsense_local/backend/modules/chat/schemas.py
@@ -35,15 +35,13 @@ class ThreadCreate(BaseModel):
 
     `source_scope` is the new chat's drafted ticks; omitted, the thread uses
     every source. `mode` is Basic (Q&A) or Agentic, kept for the thread's life;
-    omitted, the thread takes the selected model's default. `remember` makes
-    `mode` that default, and is sent only when the user chose it: a default
-    sent back as a choice would outlast the score it came from.
+    omitted, the thread takes the selected model's default. Sending a mode
+    never changes that default; the composer's switch saves a pick itself.
     """
 
     title: ThreadTitle = "New chat"
     source_scope: SourceScope | None = None
     mode: ChatMode | None = None
-    remember: bool = False
 
 
 class ThreadUpdate(BaseModel):
diff --git a/surfsense_local/backend/modules/llm/capability/modes/__init__.py b/surfsense_local/backend/modules/llm/capability/modes/__init__.py
index 2ab6a56b7d..c5047f4b1a 100644
--- a/surfsense_local/backend/modules/llm/capability/modes/__init__.py
+++ b/surfsense_local/backend/modules/llm/capability/modes/__init__.py
@@ -4,6 +4,7 @@
 from modules.llm.capability.modes.new_chat import NewChatModes, new_chat_modes
 from modules.llm.capability.modes.remembered import (
     carried_to_next_model,
+    forget_agentic,
     mode_entry,
     remember_mode,
     remembered_mode,
@@ -13,6 +14,7 @@
     "ChatMode",
     "NewChatModes",
     "carried_to_next_model",
+    "forget_agentic",
     "mode_entry",
     "new_chat_modes",
     "remember_mode",
diff --git a/surfsense_local/backend/modules/llm/capability/modes/remembered.py b/surfsense_local/backend/modules/llm/capability/modes/remembered.py
index ef1a6a1b46..ee35872e0a 100644
--- a/surfsense_local/backend/modules/llm/capability/modes/remembered.py
+++ b/surfsense_local/backend/modules/llm/capability/modes/remembered.py
@@ -8,7 +8,13 @@
 from modules.llm.capability.modes.mode import ChatMode
 from modules.llm.models import SelectedModel
 
-__all__ = ["carried_to_next_model", "mode_entry", "remember_mode", "remembered_mode"]
+__all__ = [
+    "carried_to_next_model",
+    "forget_agentic",
+    "mode_entry",
+    "remember_mode",
+    "remembered_mode",
+]
 
 KEY = "chat_modes"
 # The opt-in to try the agent the modes replaced: on, it reads as Agentic.
@@ -23,21 +29,33 @@ def remembered_mode(selected: SelectedModel) -> ChatMode | None:
     return ChatMode.AGENTIC if settings.get(_TRIAL) is True else None
 
 
-def remember_mode(
-    selected: SelectedModel, mode: ChatMode, entry: str | None = None
-) -> None:
-    """`entry` is the model the mode was chosen on, read with the choice, since the
-    slot may hold another by the time the chat opens; else the selected model."""
+def remember_mode(selected: SelectedModel, mode: ChatMode) -> None:
     # Reassigned, not mutated: a JSON column tracks assignment only.
     settings = {
         key: value
         for key, value in (selected.settings or {}).items()
         if key not in (KEY, _TRIAL)
     }
-    settings[KEY] = {**_every_mode(selected), entry or mode_entry(selected): mode.value}
+    settings[KEY] = {**_every_mode(selected), mode_entry(selected): mode.value}
     selected.settings = settings
 
 
+def forget_agentic(selected: SelectedModel, entry: str) -> None:
+    """A saved Agentic the model turned out unable to run: its own default holds again."""
+    modes = _every_mode(selected)
+    if modes.get(entry) != ChatMode.AGENTIC.value:
+        return
+    del modes[entry]
+    settings = {
+        key: value
+        for key, value in (selected.settings or {}).items()
+        if key not in (KEY, _TRIAL)
+    }
+    if modes:
+        settings[KEY] = modes
+    selected.settings = settings or None
+
+
 def carried_to_next_model(selected: SelectedModel) -> dict | None:
     """The settings the slot keeps when it takes another model: every model's mode.
 
diff --git a/surfsense_local/backend/modules/llm/capability/modes/router.py b/surfsense_local/backend/modules/llm/capability/modes/router.py
new file mode 100644
index 0000000000..e21875136e
--- /dev/null
+++ b/surfsense_local/backend/modules/llm/capability/modes/router.py
@@ -0,0 +1,49 @@
+"""The composer's mode switch: a pick is the model's default for new chats from then on."""
+
+from fastapi import APIRouter, HTTPException, status
+from pydantic import BaseModel
+from sqlalchemy.orm import Session
+
+from api.dependencies import SessionDep, transact
+from modules.llm.capability.agent_gate import catalog_facts
+from modules.llm.capability.modes.mode import ChatMode
+from modules.llm.capability.modes.new_chat import new_chat_modes
+from modules.llm.capability.modes.remembered import remember_mode
+from modules.llm.capability.read import modes_read
+from modules.llm.capability.schemas import ModesRead
+from modules.llm.model_type import ModelType
+from modules.llm.models import SelectedModel
+from modules.llm.reads_images import connection_catalog_provider
+
+__all__ = ["router"]
+
+router = APIRouter()
+
+
+class ModePick(BaseModel):
+    mode: ChatMode
+
+
+@router.put(
+    "/selection/text_gen/mode",
+    response_model=ModesRead,
+    summary="Pick the mode new chats on the chosen chat model start in",
+)
+async def pick_mode(payload: ModePick, session: SessionDep) -> ModesRead:
+    return await transact(session, _pick, payload.mode)
+
+
+def _pick(session: Session, mode: ChatMode) -> ModesRead:
+    selected = session.get(SelectedModel, ModelType.TEXT_GEN)
+    if selected is None:
+        raise HTTPException(status.HTTP_404_NOT_FOUND, "no chat model chosen")
+    facts = catalog_facts(selected, connection_catalog_provider(session, selected))
+    blocked = new_chat_modes(selected, facts).blocked
+    if mode is ChatMode.AGENTIC and blocked is not None:
+        # Never saved: the model's default stays a mode it can run.
+        raise HTTPException(
+            status.HTTP_409_CONFLICT,
+            detail={"code": blocked, "message": "This model can't run Agentic."},
+        )
+    remember_mode(selected, mode)
+    return modes_read(new_chat_modes(selected, facts))
diff --git a/surfsense_local/backend/modules/llm/capability/read.py b/surfsense_local/backend/modules/llm/capability/read.py
index b5ecd7a937..0a42dffced 100644
--- a/surfsense_local/backend/modules/llm/capability/read.py
+++ b/surfsense_local/backend/modules/llm/capability/read.py
@@ -3,7 +3,7 @@
 from modules.llm.capability.agent_gate import catalog_facts
 from modules.llm.capability.level import Level
 from modules.llm.capability.measured.loader import measured_list
-from modules.llm.capability.modes import new_chat_modes
+from modules.llm.capability.modes import NewChatModes, new_chat_modes
 from modules.llm.capability.resolve import capability_of
 from modules.llm.capability.schemas import (
     CapabilityRead,
@@ -14,7 +14,7 @@
 from modules.llm.capability.studio_output import studio_output
 from modules.llm.models import SelectedModel
 
-__all__ = ["capability_read"]
+__all__ = ["capability_read", "modes_read"]
 
 
 def capability_read(
@@ -45,12 +45,16 @@ def capability_read(
         )
         if held
         else None,
-        modes=ModesRead(
-            agentic_allowed=modes.agentic_allowed,
-            blocked=modes.blocked,
-            default_mode=modes.default_mode,
-            reason=ReasonRead(code=modes.reason.code, values=modes.reason.values),
-            remembered_mode=modes.remembered_mode,
-        ),
+        modes=modes_read(modes),
         studio=studio_output(selected),
     )
+
+
+def modes_read(modes: NewChatModes) -> ModesRead:
+    return ModesRead(
+        agentic_allowed=modes.agentic_allowed,
+        blocked=modes.blocked,
+        default_mode=modes.default_mode,
+        reason=ReasonRead(code=modes.reason.code, values=modes.reason.values),
+        remembered_mode=modes.remembered_mode,
+    )
diff --git a/surfsense_local/backend/modules/llm/router.py b/surfsense_local/backend/modules/llm/router.py
index 1760393b9c..32fae8ec2b 100644
--- a/surfsense_local/backend/modules/llm/router.py
+++ b/surfsense_local/backend/modules/llm/router.py
@@ -17,6 +17,7 @@
     model_key,
 )
 from modules.llm.capability import capability_of
+from modules.llm.capability.modes.router import router as modes_router
 from modules.llm.capability.read import capability_read
 from modules.llm.catalog.local.dependencies import LocalCatalogDep
 from modules.llm.catalog.local.engines.engine import LocalEngine
@@ -68,6 +69,7 @@
 router.include_router(voices_router)
 router.include_router(language_router)
 router.include_router(transcription_router)
+router.include_router(modes_router)
 
 
 @router.get(
diff --git a/surfsense_local/backend/tests/integration/agent/test_agent_threads.py b/surfsense_local/backend/tests/integration/agent/test_agent_threads.py
index 9dd1cbdc05..69d4d701cf 100644
--- a/surfsense_local/backend/tests/integration/agent/test_agent_threads.py
+++ b/surfsense_local/backend/tests/integration/agent/test_agent_threads.py
@@ -37,13 +37,9 @@ async def open_thread(
     api: AgentAPI,
     title: str = "New chat",
     mode: str | None = None,
-    remember: bool = False,
 ) -> dict:
-    """Open a thread the way the chat panel does, in a mode or the model's default.
-
-    `remember` is a mode the user picked in the switch, which the panel says so of.
-    """
-    fields = {"mode": mode, "remember": remember} if mode else {}
+    """Open a thread the way the chat panel does, in a mode or the model's default."""
+    fields = {"mode": mode} if mode else {}
     reply = await api.http.post(
         f"/workspaces/{api.workspace_id}/chat/threads",
         json={"title": title, **fields},
@@ -123,7 +119,11 @@ async def test_agentic_asked_for_opens_an_agent_thread_on_an_untested_model(
     """No score blocks Agentic; the user's pick is the model's default from then on."""
     monkeypatch.setattr(get_agent_settings(), "agent_untested_models", False)
 
-    chosen = await open_thread(agent_api, mode="agentic", remember=True)
+    picked = await agent_api.http.put(
+        "/llm/selection/text_gen/mode", json={"mode": "agentic"}
+    )
+    picked.raise_for_status()
+    chosen = await open_thread(agent_api, mode="agentic")
     defaulted = await open_thread(agent_api)
 
     assert chosen["uses_agent"] is True
diff --git a/surfsense_local/backend/tests/integration/agent/test_engine_choice.py b/surfsense_local/backend/tests/integration/agent/test_engine_choice.py
index 89bf05ebfe..96fbf181dc 100644
--- a/surfsense_local/backend/tests/integration/agent/test_engine_choice.py
+++ b/surfsense_local/backend/tests/integration/agent/test_engine_choice.py
@@ -14,10 +14,9 @@
 from modules.agent.engine_choice import (
     AgenticRefusedError,
     new_thread_mode,
-    remember_thread_mode,
     selected_model_can_run_agent,
 )
-from modules.llm.capability.modes import ChatMode, remembered_mode
+from modules.llm.capability.modes import ChatMode, remember_mode, remembered_mode
 from modules.llm.model_type import ModelType
 from modules.llm.models import ProviderConnection, SelectedModel
 from shared.config import get_agent_settings, get_llm_settings
@@ -275,6 +274,25 @@ async def test_a_local_model_whose_template_parses_no_tool_calls_is_refused(
     assert (await new_thread_mode(session, None)).mode is ChatMode.BASIC
 
 
+async def test_agentic_saved_for_a_local_model_that_cannot_run_it_is_dropped_once_refused(
+    session: Session, no_switch: None, local_runtime: LocalRuntime
+) -> None:
+    """A local model's tool calls are read only when a chat starts, so the pick
+    could be saved; the refusal puts the model back on its own default."""
+    local_runtime.template_caps = {"supports_tools": True, "supports_tool_calls": False}
+    select_local(session)
+    selected = session.get(SelectedModel, ModelType.TEXT_GEN)
+    assert selected is not None
+    remember_mode(selected, ChatMode.AGENTIC)
+    session.commit()
+
+    assert await refused(session, ChatMode.AGENTIC) == "tool_calls_unsupported"
+
+    session.expire_all()
+    selected = session.get(SelectedModel, ModelType.TEXT_GEN)
+    assert selected is not None and remembered_mode(selected) is None
+
+
 async def test_a_local_model_loaded_with_a_short_window_is_refused(
     session: Session, no_switch: None, local_runtime: LocalRuntime
 ) -> None:
@@ -305,22 +323,6 @@ async def test_a_local_copy_of_a_model_that_passed_opens_basic(
     assert (await new_thread_mode(session, None)).mode is ChatMode.BASIC
 
 
-async def test_the_mode_a_chat_opened_in_is_the_model_s_next_default(
-    session: Session, no_switch: None
-) -> None:
-    """Remembered on the selection, over the measured default."""
-    select_remote(session, "gpt-4o-mini", "openai", base_url=OPENAI)
-    opening = await new_thread_mode(session, ChatMode.AGENTIC)
-    assert opening.entry is not None
-
-    await remember_thread_mode(session, opening.entry, ChatMode.AGENTIC)
-
-    assert (await new_thread_mode(session, None)).mode is ChatMode.AGENTIC
-    session.expire_all()
-    selected = session.get(SelectedModel, ModelType.TEXT_GEN)
-    assert selected is not None and remembered_mode(selected) is ChatMode.AGENTIC
-
-
 async def test_an_agent_thread_s_next_turn_is_held_only_to_technical_gates(
     session: Session, no_switch: None
 ) -> None:
diff --git a/surfsense_local/backend/tests/integration/agent/test_thread_modes.py b/surfsense_local/backend/tests/integration/agent/test_thread_modes.py
index 88e4a8a243..83697c5f82 100644
--- a/surfsense_local/backend/tests/integration/agent/test_thread_modes.py
+++ b/surfsense_local/backend/tests/integration/agent/test_thread_modes.py
@@ -1,6 +1,4 @@
-"""Opening a chat in Basic (Q&A) or Agentic through the API: refusals leave no thread, and the choice is remembered."""
-
-import asyncio
+"""Opening a chat in Basic (Q&A) or Agentic through the API: refusals leave no thread, and opening one saves no default."""
 
 import pytest
 from httpx import AsyncClient
@@ -8,7 +6,6 @@
 
 from modules.agent.engines import opencode_engine
 from modules.chat.models import ChatThread
-from modules.llm.capability.modes import carried_to_next_model
 from modules.llm.model_type import ModelType
 from modules.llm.models import ProviderConnection, SelectedModel
 from shared.config import get_agent_settings
@@ -70,56 +67,10 @@ async def _open(client: AsyncClient, **fields: object) -> tuple[int, dict]:
     return reply.status_code, reply.json()
 
 
-async def test_basic_chosen_opens_a_chat_and_is_remembered(
-    client: AsyncClient, engine: Engine
-) -> None:
-    """The user's pick becomes the model's default for new chats."""
-    _select(engine, "gpt-4o-mini")
-
-    status, thread = await _open(client, mode="basic", remember=True)
-
-    assert status == 201
-    assert thread["uses_agent"] is False
-    capability = (await client.get("/llm/selection/text_gen")).json()["capability"]
-    assert capability["modes"]["remembered_mode"] == "basic"
-
-
-async def test_a_pick_is_remembered_for_the_model_it_was_made_on(
-    client: AsyncClient, engine: Engine, monkeypatch: pytest.MonkeyPatch
-) -> None:
-    """The agent may take a minute to start, and the model picker stays open meanwhile."""
-    connection_id = _select(engine, "gpt-4o-mini")
-
-    def select_another() -> None:
-        # As the selection route leaves the slot: another model, every mode carried.
-        with create_session_factory(engine)() as session:
-            selected = session.get(SelectedModel, ModelType.TEXT_GEN)
-            assert selected is not None
-            selected.settings = carried_to_next_model(selected)
-            selected.name = "gpt-4.1"
-            session.commit()
-
-    async def another_model_while_starting(*_args: object) -> str:
-        await asyncio.to_thread(select_another)
-        return "ses_1"
-
-    monkeypatch.setattr(
-        opencode_engine, "open_agent_session", another_model_while_starting
-    )
-
-    status, thread = await _open(client, mode="agentic", remember=True)
-
-    assert status == 201
-    assert thread["uses_agent"] is True
-    assert _remembered(engine) == {
-        "chat_modes": {f"openai_compatible/{connection_id}/gpt-4o-mini": "agentic"}
-    }
-
-
-async def test_a_default_sent_back_unchosen_is_not_remembered(
+async def test_a_mode_sent_with_a_new_chat_is_not_saved_as_the_default(
     client: AsyncClient, engine: Engine
 ) -> None:
-    """A switch the user never touched: a later score, or the developer switch
+    """Only the switch saves a pick: a later score, or the developer switch
     turned off, still decides the next chat."""
     _select(engine, "gpt-4o-mini")
 
@@ -138,7 +89,7 @@ async def test_agentic_refused_by_a_gate_is_a_409_with_its_code_and_no_thread(
     """The catalog says gpt-3.5-turbo makes no tool calls."""
     _select(engine, "gpt-3.5-turbo")
 
-    status, body = await _open(client, mode="agentic", remember=True)
+    status, body = await _open(client, mode="agentic")
 
     assert status == 409
     assert body["detail"]["code"] == "tool_calls_unsupported"
@@ -157,7 +108,7 @@ async def not_ready(*_args: object) -> None:
     monkeypatch.setattr(opencode_engine, "open_agent_session", not_ready)
     _select(engine, "gpt-4o-mini")
 
-    status, body = await _open(client, mode="agentic", remember=True)
+    status, body = await _open(client, mode="agentic")
 
     assert status == 503
     assert body["detail"]["code"] == "agent_unavailable"
diff --git a/surfsense_local/backend/tests/integration/llm/test_capability_routes.py b/surfsense_local/backend/tests/integration/llm/test_capability_routes.py
index 3a5689a628..7d36663516 100644
--- a/surfsense_local/backend/tests/integration/llm/test_capability_routes.py
+++ b/surfsense_local/backend/tests/integration/llm/test_capability_routes.py
@@ -227,12 +227,8 @@ async def select(name: str) -> dict:
         return chosen.json()["capability"]["modes"]
 
     await select("Qwen3-4B-Q4_K_M")
-    workspace = (await client.post("/workspaces", json={"name": "Research"})).json()
-    opened = await client.post(
-        f"/workspaces/{workspace['id']}/chat/threads",
-        json={"title": "New chat", "mode": "basic", "remember": True},
-    )
-    assert opened.status_code == 201, opened.text
+    picked = await client.put("/llm/selection/text_gen/mode", json={"mode": "basic"})
+    assert picked.status_code == 200, picked.text
 
     other = await select("Qwen3-1.7B-Q4_K_M")
     back = await select("Qwen3-4B-Q4_K_M")
diff --git a/surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py b/surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py
new file mode 100644
index 0000000000..bb5f413e97
--- /dev/null
+++ b/surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py
@@ -0,0 +1,79 @@
+"""Picking Basic (Q&A) or Agentic in the composer saves it as the model's default at once."""
+
+import pytest
+from httpx import AsyncClient
+from sqlalchemy import Engine
+
+from modules.llm.model_type import ModelType
+from modules.llm.models import ProviderConnection, SelectedModel
+from shared.config import get_agent_settings
+from shared.db import create_session_factory
+
+pytestmark = pytest.mark.integration
+
+OPENROUTER = "https://openrouter.ai/api/v1"
+OPENAI = "https://api.openai.com/v1"
+
+
+@pytest.fixture(autouse=True)
+def opencode_staged(monkeypatch: pytest.MonkeyPatch) -> None:
+    """As an installer ships: opencode beside the API, no developer switch."""
+    settings = get_agent_settings()
+    monkeypatch.setattr(settings, "opencode_url", "http://127.0.0.1:9")
+    monkeypatch.setattr(settings, "opencode_password", "not-a-running-opencode")
+    monkeypatch.setattr(settings, "agent_untested_models", False)
+
+
+def _select(engine: Engine, name: str, catalog_provider: str, base_url: str) -> None:
+    """A remote selection written as the selection route leaves it; nothing is sent to the host."""
+    with create_session_factory(engine)() as session:
+        connection = ProviderConnection(
+            label=f"Remote {name}",
+            provider="openai_compatible",
+            base_url=base_url,
+            catalog_provider=catalog_provider,
+        )
+        session.add(connection)
+        session.flush()
+        session.merge(
+            SelectedModel(
+                model_type=ModelType.TEXT_GEN,
+                provider="openai_compatible",
+                connection_id=connection.id,
+                name=name,
+            )
+        )
+        session.commit()
+
+
+async def _modes(client: AsyncClient) -> dict:
+    reply = await client.get("/llm/selection/text_gen")
+    assert reply.status_code == 200, reply.text
+    return reply.json()["capability"]["modes"]
+
+
+async def test_basic_picked_on_an_agentic_model_is_its_default_without_a_chat(
+    client: AsyncClient, engine: Engine
+) -> None:
+    """A refresh before the first message keeps the pick."""
+    _select(engine, "anthropic/claude-haiku-4.5", "openrouter", OPENROUTER)
+
+    picked = await client.put("/llm/selection/text_gen/mode", json={"mode": "basic"})
+
+    assert picked.status_code == 200, picked.text
+    modes = await _modes(client)
+    assert (modes["default_mode"], modes["remembered_mode"]) == ("basic", "basic")
+
+
+async def test_agentic_picked_on_a_model_that_cannot_run_it_is_refused_and_not_kept(
+    client: AsyncClient, engine: Engine
+) -> None:
+    """The catalog says gpt-3.5-turbo makes no tool calls."""
+    _select(engine, "gpt-3.5-turbo", "openai", OPENAI)
+
+    picked = await client.put("/llm/selection/text_gen/mode", json={"mode": "agentic"})
+
+    assert picked.status_code == 409, picked.text
+    assert picked.json()["detail"]["code"] == "tool_calls_unsupported"
+    modes = await _modes(client)
+    assert (modes["default_mode"], modes["remembered_mode"]) == ("basic", None)
diff --git a/surfsense_local/backend/tests/unit/llm/capability/test_new_chat_modes.py b/surfsense_local/backend/tests/unit/llm/capability/test_new_chat_modes.py
index 50415c27cd..d26e83f37e 100644
--- a/surfsense_local/backend/tests/unit/llm/capability/test_new_chat_modes.py
+++ b/surfsense_local/backend/tests/unit/llm/capability/test_new_chat_modes.py
@@ -8,7 +8,6 @@
 from modules.llm.capability.modes import (
     ChatMode,
     carried_to_next_model,
-    mode_entry,
     new_chat_modes,
     remember_mode,
     remembered_mode,
@@ -279,19 +278,6 @@ def test_each_model_keeps_its_own_mode_when_the_slot_takes_another() -> None:
     assert after_trial.settings is not None and "voices" not in after_trial.settings
 
 
-def test_a_mode_chosen_on_a_model_the_slot_has_left_is_that_model_s() -> None:
-    """The slot took another model while the chat opened: the new one's old opt-in stays."""
-    chosen_on = _selected("gpt-4o-mini")
-    now = _selected("gpt-4.1", settings={"agent_trial": True})
-
-    remember_mode(now, ChatMode.BASIC, mode_entry(chosen_on))
-
-    assert remembered_mode(now) is ChatMode.AGENTIC
-    assert remembered_mode(_selected("gpt-4o-mini", settings=now.settings)) is (
-        ChatMode.BASIC
-    )
-
-
 def test_the_same_name_on_another_server_is_another_model() -> None:
     """A local copy and a provider's host each keep their own choice."""
     remote = _selected("qwen/qwen3.8-27b", remembered=ChatMode.AGENTIC)
diff --git a/surfsense_local/frontend/src/features/chat/api.ts b/surfsense_local/frontend/src/features/chat/api.ts
index 8c431e9fa9..f94ca5d3cf 100644
--- a/surfsense_local/frontend/src/features/chat/api.ts
+++ b/surfsense_local/frontend/src/features/chat/api.ts
@@ -111,8 +111,6 @@ export function createThread(
     body: JSON.stringify({
       title,
       ...(choice ? { mode: choice.mode } : {}),
-      // Only the user's own choice becomes the model's default.
-      ...(choice?.chosen ? { remember: true } : {}),
     }),
     signal,
   })
diff --git a/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx b/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
index 55167f5a52..9ee72ad4d0 100644
--- a/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
+++ b/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
@@ -131,6 +131,35 @@ describe("the composer's mode switch", () => {
     expect(composerNote()).toEqual([])
   })
 
+  it("saves a pick as the model's default, so a refresh keeps it", async () => {
+    const user = userEvent.setup()
+    renderComposer({ model: KIMI })
+    const fetchMock = vi.mocked(fetch)
+    fetchMock.mockImplementation(async () =>
+      Response.json({
+        ...PASSED,
+        default_mode: "basic",
+        remembered_mode: "basic",
+      })
+    )
+
+    await user.click(trigger())
+    await user.click(
+      await screen.findByRole("menuitemradio", { name: /^Basic \(Q&A\)/ })
+    )
+
+    await waitFor(() =>
+      expect(fetchMock).toHaveBeenCalledWith(
+        expect.stringContaining("/llm/selection/text_gen/mode"),
+        expect.objectContaining({
+          method: "PUT",
+          body: JSON.stringify({ mode: "basic" }),
+        })
+      )
+    )
+    expect(nextChat()).toBe("basic")
+  })
+
   it("starts an untested model in Basic, and warns only in the menu", async () => {
     const user = userEvent.setup()
     renderComposer({ model: SONNET_99 })
diff --git a/surfsense_local/frontend/src/features/chat/modes/mode-picker.tsx b/surfsense_local/frontend/src/features/chat/modes/mode-picker.tsx
index ae9dff2def..8f17f77b92 100644
--- a/surfsense_local/frontend/src/features/chat/modes/mode-picker.tsx
+++ b/surfsense_local/frontend/src/features/chat/modes/mode-picker.tsx
@@ -26,11 +26,8 @@ import {
   modeLabel,
   modeShortDescription,
 } from "./mode-text"
-import {
-  useChatModes,
-  useNewChatMode,
-  usePickNewChatMode,
-} from "./new-chat-mode"
+import { useChatModes, useNewChatMode } from "./new-chat-mode"
+import { useSaveModePick } from "./save-mode-pick"
 
 const MODE_ICON = { basic: Chat01Icon, agentic: BotIcon } as const
 
@@ -52,7 +49,7 @@ export function ModePicker({
   className?: string
 }) {
   const newChatMode = useNewChatMode(model)
-  const pick = usePickNewChatMode()
+  const savePick = useSaveModePick()
   const modes = useChatModes(model)
   if (!modes || !newChatMode) return null
 
@@ -81,7 +78,7 @@ export function ModePicker({
             <ModeList
               modes={modes}
               value={newChatMode}
-              onPick={(mode) => pick(model, mode)}
+              onPick={(mode) => savePick(model, mode)}
             />
           </DropdownMenuGroup>
         </DropdownMenuContent>
@@ -117,7 +114,7 @@ export function ModePicker({
             // A chat keeps its mode, so the other one opens a new chat.
             onPick={(mode) => {
               if (mode === threadMode) return
-              pick(model, mode)
+              savePick(model, mode)
               onNewChat?.()
             }}
             otherOpensNewChat
diff --git a/surfsense_local/frontend/src/features/chat/modes/new-chat-mode.ts b/surfsense_local/frontend/src/features/chat/modes/new-chat-mode.ts
index 557277291e..ece38d92ef 100644
--- a/surfsense_local/frontend/src/features/chat/modes/new-chat-mode.ts
+++ b/surfsense_local/frontend/src/features/chat/modes/new-chat-mode.ts
@@ -8,7 +8,7 @@ import {
 } from "@/features/models/selection/api"
 
 export type NewChatModePicks = {
-  // Per model, the mode picked this session; the API remembers it once a chat opens in it.
+  // Per model, the mode picked this session, shown before the API has saved it.
   picks: Readonly<Record<string, ChatMode>>
   pick: (model: SelectionTarget, mode: ChatMode) => void
   // Per model, the gate the API refused an Agentic chat with this session: a
@@ -17,11 +17,8 @@ export type NewChatModePicks = {
   block: (model: SelectionTarget, gate: string) => void
 }
 
-/**
- * The mode the next new chat opens in, and whether the user chose it. Only a
- * choice is remembered: a default sent back as one would outlast its score.
- */
-export type NewChatChoice = { mode: ChatMode; chosen: boolean }
+/** The mode the next new chat opens in. */
+export type NewChatChoice = { mode: ChatMode }
 
 // Without a provider, nothing is picked and the model's own default holds.
 export const NewChatModeContext = createContext<NewChatModePicks>({
@@ -65,7 +62,7 @@ export function resolveNewChatChoice(
   const picked = picks[modelKey(model)]
   const wanted = picked ?? modes.default_mode
   const mode = wanted === "agentic" && !modes.agentic_allowed ? "basic" : wanted
-  return { mode, chosen: picked === mode }
+  return { mode }
 }
 
 export function useChatModes(model: ModelSelection | null) {
@@ -77,12 +74,8 @@ export function useNewChatChoice(model: ModelSelection | null) {
   const { picks, blocks } = useContext(NewChatModeContext)
   const choice = resolveNewChatChoice(model, picks, blocks)
   const mode = choice?.mode ?? null
-  const chosen = choice?.chosen ?? false
   // Held while it says the same, so the runtime's send is not made anew.
-  return useMemo<NewChatChoice | null>(
-    () => (mode ? { mode, chosen } : null),
-    [mode, chosen]
-  )
+  return useMemo<NewChatChoice | null>(() => (mode ? { mode } : null), [mode])
 }
 
 export function useNewChatMode(model: ModelSelection | null) {
diff --git a/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts b/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
new file mode 100644
index 0000000000..bc1928179c
--- /dev/null
+++ b/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
@@ -0,0 +1,52 @@
+import { useQueryClient } from "@tanstack/react-query"
+import { useCallback } from "react"
+import { toast } from "sonner"
+
+import type { ChatMode, ChatModes } from "@/features/models/capability/api"
+import { modelKey, type ModelSelection } from "@/features/models/selection/api"
+import { selectionQueryKey } from "@/features/models/selection/use-selection"
+import { intl } from "@/i18n/intl"
+import { requestJson } from "@/lib/api"
+
+import { usePickNewChatMode } from "./new-chat-mode"
+
+function saveModePick(mode: ChatMode): Promise<ChatModes> {
+  return requestJson<ChatModes>("/llm/selection/text_gen/mode", {
+    method: "PUT",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ mode }),
+  })
+}
+
+/**
+ * A pick shows at once and is saved as the model's default, so a refresh or a
+ * restart keeps it. Unsaved, it still holds for the session.
+ */
+export function useSaveModePick() {
+  const pick = usePickNewChatMode()
+  const client = useQueryClient()
+  return useCallback(
+    (model: ModelSelection, mode: ChatMode) => {
+      pick(model, mode)
+      saveModePick(mode).then(
+        (modes) =>
+          client.setQueryData<ModelSelection | null>(
+            selectionQueryKey("text_gen"),
+            (current) =>
+              current?.capability && modelKey(current) === modelKey(model)
+                ? { ...current, capability: { ...current.capability, modes } }
+                : current
+          ),
+        () =>
+          toast.error(
+            intl.formatMessage({
+              id: "chat_mode_picker_not_saved_toast",
+              defaultMessage:
+                "Couldn’t save this mode. New chats use it until you restart.",
+            })
+          )
+      )
+    },
+    [pick, client]
+  )
+}
diff --git a/surfsense_local/frontend/src/features/chat/use-chat-runtime.test.tsx b/surfsense_local/frontend/src/features/chat/use-chat-runtime.test.tsx
index 838d89a243..54d136ef13 100644
--- a/surfsense_local/frontend/src/features/chat/use-chat-runtime.test.tsx
+++ b/surfsense_local/frontend/src/features/chat/use-chat-runtime.test.tsx
@@ -986,23 +986,9 @@ describe("while a reply streams", () => {
 })
 
 describe("a new chat's mode", () => {
-  it("opens the chat in the mode picked for the model, and has it remembered", async () => {
+  it("opens the chat in the mode the switch shows", async () => {
     const { result } = renderRuntime({
-      newChatMode: { mode: "agentic", chosen: true },
-    })
-    await waitFor(() => expect(result.current.threads.length).toBe(2))
-
-    sendIn(result, "Draft the board pack")
-
-    await waitFor(() => expect(api.sendStreams).toHaveLength(1))
-    expect(api.creates).toEqual([
-      { title: "New chat", mode: "agentic", remember: true },
-    ])
-  })
-
-  it("opens the chat in the model's default without remembering it", async () => {
-    const { result } = renderRuntime({
-      newChatMode: { mode: "basic", chosen: false },
+      newChatMode: { mode: "basic" },
     })
     await waitFor(() => expect(result.current.threads.length).toBe(2))
 
@@ -1027,7 +1013,7 @@ describe("a new chat's mode", () => {
     onTestFinished(() => shown.mockRestore())
     api.refuseCreate = { status: 409, code: "tool_calls_unsupported" }
     const { result } = renderRuntime({
-      newChatMode: { mode: "agentic", chosen: true },
+      newChatMode: { mode: "agentic" },
     })
     await waitFor(() => expect(result.current.threads.length).toBe(2))
 
diff --git a/surfsense_local/frontend/src/features/dashboard/dashboard-page.test.tsx b/surfsense_local/frontend/src/features/dashboard/dashboard-page.test.tsx
index 8eb70e44bb..f4258cb80f 100644
--- a/surfsense_local/frontend/src/features/dashboard/dashboard-page.test.tsx
+++ b/surfsense_local/frontend/src/features/dashboard/dashboard-page.test.tsx
@@ -1110,10 +1110,8 @@ describe("dashboard chat", () => {
     )
 
     expect(await within(conversation).findByText("Revenue rose.")).toBeTruthy()
-    // Picked in the switch, so it becomes the model's default.
-    expect(created).toEqual([
-      { title: "New chat", mode: "basic", remember: true },
-    ])
+    // The switch saved the pick itself; the new chat only says its mode.
+    expect(created).toEqual([{ title: "New chat", mode: "basic" }])
     await user.click(
       within(conversation).getByRole("button", { name: /^Chat mode Basic/ })
     )
diff --git a/surfsense_local/frontend/translations/de.json b/surfsense_local/frontend/translations/de.json
index 9e01b4ff35..c2e7fa5e27 100644
--- a/surfsense_local/frontend/translations/de.json
+++ b/surfsense_local/frontend/translations/de.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "Antwortet aus deinen Quellen",
   "chat_mode_picker_change_tooltip": "Basis (Q&A) oder Agentisch",
   "chat_mode_picker_new_chat_label": "Modus für diesen Chat",
+  "chat_mode_picker_not_saved_toast": "Der Modus konnte nicht gespeichert werden. Neue Chats nutzen ihn bis zum Neustart.",
   "chat_mode_picker_open_chat_label": "Modus dieses Chats",
   "chat_mode_picker_opens_new_chat_label": "Neuer Chat",
   "chat_mode_picker_thread_aria": "Chatmodus {mode}. Ein Chat behält seinen Modus; starte für einen anderen einen neuen Chat.",
diff --git a/surfsense_local/frontend/translations/en.json b/surfsense_local/frontend/translations/en.json
index b6faad23d4..49df3b5b8f 100644
--- a/surfsense_local/frontend/translations/en.json
+++ b/surfsense_local/frontend/translations/en.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "Answers from your sources",
   "chat_mode_picker_change_tooltip": "Basic (Q&A) or Agentic",
   "chat_mode_picker_new_chat_label": "Mode for this chat",
+  "chat_mode_picker_not_saved_toast": "Couldn’t save this mode. New chats use it until you restart.",
   "chat_mode_picker_open_chat_label": "This chat’s mode",
   "chat_mode_picker_opens_new_chat_label": "New chat",
   "chat_mode_picker_thread_aria": "Chat mode {mode}. A chat keeps its mode; start a new chat for another.",
diff --git a/surfsense_local/frontend/translations/es.json b/surfsense_local/frontend/translations/es.json
index c1892879f2..76b13cdc2e 100644
--- a/surfsense_local/frontend/translations/es.json
+++ b/surfsense_local/frontend/translations/es.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "Responde a partir de tus fuentes",
   "chat_mode_picker_change_tooltip": "Básico (P&R) o Agéntico",
   "chat_mode_picker_new_chat_label": "Modo de este chat",
+  "chat_mode_picker_not_saved_toast": "No se pudo guardar este modo. Los chats nuevos lo usarán hasta que reinicies.",
   "chat_mode_picker_open_chat_label": "Modo de este chat",
   "chat_mode_picker_opens_new_chat_label": "Nuevo chat",
   "chat_mode_picker_thread_aria": "Modo de chat {mode}. Un chat conserva su modo; inicia un chat nuevo para usar otro.",
diff --git a/surfsense_local/frontend/translations/fr.json b/surfsense_local/frontend/translations/fr.json
index f5aaec71f6..d16436156d 100644
--- a/surfsense_local/frontend/translations/fr.json
+++ b/surfsense_local/frontend/translations/fr.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "Répond à partir de vos sources",
   "chat_mode_picker_change_tooltip": "Basique (Q&R) ou Agentique",
   "chat_mode_picker_new_chat_label": "Mode de ce chat",
+  "chat_mode_picker_not_saved_toast": "Impossible d’enregistrer ce mode. Les nouveaux chats l’utilisent jusqu’au redémarrage.",
   "chat_mode_picker_open_chat_label": "Mode de ce chat",
   "chat_mode_picker_opens_new_chat_label": "Nouveau chat",
   "chat_mode_picker_thread_aria": "Mode de chat {mode}. Un chat garde son mode ; démarrez un nouveau chat pour en changer.",
diff --git a/surfsense_local/frontend/translations/hi.json b/surfsense_local/frontend/translations/hi.json
index 7584b7f2b3..d06d73ccaf 100644
--- a/surfsense_local/frontend/translations/hi.json
+++ b/surfsense_local/frontend/translations/hi.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "आपके स्रोतों से जवाब देता है",
   "chat_mode_picker_change_tooltip": "बेसिक (सवाल-जवाब) या एजेंटिक",
   "chat_mode_picker_new_chat_label": "इस चैट का मोड",
+  "chat_mode_picker_not_saved_toast": "यह मोड सहेजा नहीं जा सका। रीस्टार्ट करने तक नई चैट इसी मोड में खुलेंगी।",
   "chat_mode_picker_open_chat_label": "इस चैट का मोड",
   "chat_mode_picker_opens_new_chat_label": "नई चैट",
   "chat_mode_picker_thread_aria": "चैट मोड {mode}. चैट अपना मोड बनाए रखती है; दूसरे मोड के लिए नई चैट शुरू करें।",
diff --git a/surfsense_local/frontend/translations/ja.json b/surfsense_local/frontend/translations/ja.json
index e9c21b7c7b..296a6e0029 100644
--- a/surfsense_local/frontend/translations/ja.json
+++ b/surfsense_local/frontend/translations/ja.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "ソースをもとに回答",
   "chat_mode_picker_change_tooltip": "ベーシック（Q&A）またはエージェント",
   "chat_mode_picker_new_chat_label": "このチャットのモード",
+  "chat_mode_picker_not_saved_toast": "このモードを保存できませんでした。再起動するまで、新しいチャットはこのモードで開きます。",
   "chat_mode_picker_open_chat_label": "このチャットのモード",
   "chat_mode_picker_opens_new_chat_label": "新しいチャット",
   "chat_mode_picker_thread_aria": "チャットモード: {mode}。チャットのモードは変わりません。別のモードにするには新しいチャットを始めてください。",
diff --git a/surfsense_local/frontend/translations/ko.json b/surfsense_local/frontend/translations/ko.json
index 858ecd3027..3d14d6c955 100644
--- a/surfsense_local/frontend/translations/ko.json
+++ b/surfsense_local/frontend/translations/ko.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "소스를 바탕으로 답변",
   "chat_mode_picker_change_tooltip": "기본(Q&A) 또는 에이전트",
   "chat_mode_picker_new_chat_label": "이 채팅의 모드",
+  "chat_mode_picker_not_saved_toast": "이 모드를 저장하지 못했습니다. 다시 시작할 때까지 새 채팅은 이 모드로 열립니다.",
   "chat_mode_picker_open_chat_label": "이 채팅의 모드",
   "chat_mode_picker_opens_new_chat_label": "새 채팅",
   "chat_mode_picker_thread_aria": "채팅 모드 {mode}. 채팅은 모드를 유지합니다. 다른 모드를 쓰려면 새 채팅을 시작하세요.",
diff --git a/surfsense_local/frontend/translations/pt-BR.json b/surfsense_local/frontend/translations/pt-BR.json
index e2ca00d053..52674766b2 100644
--- a/surfsense_local/frontend/translations/pt-BR.json
+++ b/surfsense_local/frontend/translations/pt-BR.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "Responde com base nas suas fontes",
   "chat_mode_picker_change_tooltip": "Básico (P&R) ou Agêntico",
   "chat_mode_picker_new_chat_label": "Modo deste chat",
+  "chat_mode_picker_not_saved_toast": "Não foi possível salvar este modo. Novos chats vão usá-lo até você reiniciar.",
   "chat_mode_picker_open_chat_label": "Modo deste chat",
   "chat_mode_picker_opens_new_chat_label": "Novo chat",
   "chat_mode_picker_thread_aria": "Modo do chat {mode}. Um chat mantém o modo; inicie um novo chat para usar outro.",
diff --git a/surfsense_local/frontend/translations/ru.json b/surfsense_local/frontend/translations/ru.json
index 3314f17637..b1c221c2fd 100644
--- a/surfsense_local/frontend/translations/ru.json
+++ b/surfsense_local/frontend/translations/ru.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "Отвечает по вашим источникам",
   "chat_mode_picker_change_tooltip": "Базовый (вопрос-ответ) или агентный",
   "chat_mode_picker_new_chat_label": "Режим этого чата",
+  "chat_mode_picker_not_saved_toast": "Не удалось сохранить этот режим. Новые чаты будут открываться в нём до перезапуска.",
   "chat_mode_picker_open_chat_label": "Режим этого чата",
   "chat_mode_picker_opens_new_chat_label": "Новый чат",
   "chat_mode_picker_thread_aria": "Режим чата: {mode}. Чат сохраняет свой режим; для другого начните новый чат.",
diff --git a/surfsense_local/frontend/translations/zh-CN.json b/surfsense_local/frontend/translations/zh-CN.json
index 0ed6bcc470..67a778bbd1 100644
--- a/surfsense_local/frontend/translations/zh-CN.json
+++ b/surfsense_local/frontend/translations/zh-CN.json
@@ -214,6 +214,7 @@
   "chat_mode_basic_short_body": "根据你的来源回答",
   "chat_mode_picker_change_tooltip": "基础（问答）或智能体",
   "chat_mode_picker_new_chat_label": "此对话的模式",
+  "chat_mode_picker_not_saved_toast": "无法保存此模式。重启前，新对话仍会使用它。",
   "chat_mode_picker_open_chat_label": "此对话的模式",
   "chat_mode_picker_opens_new_chat_label": "新对话",
   "chat_mode_picker_thread_aria": "对话模式 {mode}。对话会保持其模式；要换模式，请开始新对话。",

From 87b8edea80da99a6016c1acc187917d71219d198 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 12:34:43 +0530
Subject: [PATCH 06/12] feat(local): fade plugin icons in once they load

Each logo popped in on its own, which made a long list jump. The tile now shows
at once and the logo fades in on it; one the browser already holds does not fade.
---
 docs/architecture/plugins.md                  |  2 +-
 .../src/features/plugins/plugin-icon.test.tsx | 47 +++++++++++++++++++
 .../src/features/plugins/plugin-icon.tsx      | 30 ++++++++----
 3 files changed, 70 insertions(+), 9 deletions(-)
 create mode 100644 surfsense_local/frontend/src/features/plugins/plugin-icon.test.tsx

diff --git a/docs/architecture/plugins.md b/docs/architecture/plugins.md
index 47489c75c2..cc01533f70 100644
--- a/docs/architecture/plugins.md
+++ b/docs/architecture/plugins.md
@@ -137,7 +137,7 @@ A plugin's server gets the arguments of the calls opencode makes to it, and noth
 - **Search**, beside the tabs, finds plugins in both at once ([`search/`](../../surfsense_local/frontend/src/features/plugins/search/)): a plugin matches when every word of the query is in its name, id, description, author or tool names, or, for a connected one, its host and tool titles. Results list **Yours**, the connected plugins in the same table, then **Available**, the first-party and registry plugins not connected, as one-line rows rather than Browse's cards: icon, name and description, with Connect; the registry is searched only while third-party plugins are allowed, and the results say so otherwise. While the query is not empty the tabs are held and the results replace them; Escape or the clear button gives them back. Results settle 200 ms after the last keystroke, with a skeleton of both sections until then. A result opens its details or its public page in place, and Back returns to the results.
 - **Yours** ([`your-plugins.tsx`](../../surfsense_local/frontend/src/features/plugins/yours/your-plugins.tsx)) is a table of connected plugins, Plugin, Type and Status ([`connected-plugins-table.tsx`](../../surfsense_local/frontend/src/features/plugins/yours/connected-plugins-table.tsx)). Type is "By SurfSense", "Third-party" or "Custom" for a server added by its URL. A third-party plugin while third-party plugins are off says "Third-party plugins are off" before anything else, with no Connect. Otherwise a plugin not connected, of any type, shows **Connect** in place of its status, the same Connect as its details ([`not-connected.ts`](../../surfsense_local/frontend/src/features/plugins/not-connected.ts), [`connect-plugin-button.tsx`](../../surfsense_local/frontend/src/features/plugins/details/connect-plugin-button.tsx)): disconnected, or lacking a sign-in, key, headers or OAuth app. Beside it, in grey, is why, so the user knows what Connect will ask: "Needs an OAuth app", "Needs more access", "Sign-in expired" or "Key refused"; nothing for a plugin disconnected by the user or never signed in ([`connect-reason.ts`](../../surfsense_local/frontend/src/features/plugins/yours/connect-reason.ts)). An OAuth plugin's Connect opens its details, where the sign-in runs. Any other status is one state, the first that holds, red when something broke, amber when it needs a step from the user, grey when off on purpose ([`plugin-status.ts`](../../surfsense_local/frontend/src/features/plugins/yours/plugin-status.ts), [`status-tone.ts`](../../surfsense_local/frontend/src/features/plugins/yours/status-tone.ts)): "No longer listed", a failed check connecting does not fix: in red "Couldn’t connect", "License revoked" or "License not recognized"; in amber "Needs a license" or "License expired", as "No longer listed" is, "Off", then "On". "On" carries a "New tools" badge while a tool is new or changed. Each row opens the plugin's details. With nothing connected it says so, with **Browse plugins** and **Add custom plugin** ([`no-plugins-connected.tsx`](../../surfsense_local/frontend/src/features/plugins/yours/no-plugins-connected.tsx)).
 - **Browse** ([`browse-plugins.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/browse-plugins.tsx)) holds **By SurfSense**, from the first-party list ([`first-party-plugins.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/first-party-plugins.tsx)): each plugin with its logo, its description, "License required" where `access` is `license`, and **Connect**, or, once saved, its status as Yours shows it and in the same colour, which opens its details there ([`saved-plugin-status.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/saved-plugin-status.tsx)); connecting ends in a toast wherever it starts, Browse, Add custom plugin, Connect on a disconnected plugin in Yours, a key or a sign-in: "{plugin} connected", including once an account's sign-in finishes in the browser, or "Couldn’t connect {plugin}" with the reason in the app's words and Report issue ([`connect-toasts.ts`](../../surfsense_local/frontend/src/features/plugins/connect-toasts.ts)). Then **Third-party** ([`third-party-plugins.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/third-party-plugins.tsx)), its header holding **Allow third-party plugins**, off by default. Turning it on first asks, saying other people run those servers, what a tool is sent goes to its publisher, and SurfSense does not review changes to them; turning it off asks nothing. While it is off nothing offers to reach a third-party server: no Connect in Yours or on a plugin's details, and Add custom plugin holds Connect, saying a server added by URL is a third-party plugin, beside the same switch and consent ([`third-party-held.ts`](../../surfsense_local/frontend/src/features/plugins/third-party-held.ts)). While it is off the section is one sentence; while it is on it holds the registry's list once they are allowed ([`registry-plugins.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/registry-plugins.tsx)), dated "Updated" with `fetched_at`. While `plugins.surfsense.com` is off it says the list comes from there, with **Allow**, which raises the egress prompt on the click rather than at load. Each entry shows its icon or its name's first letter ([`plugin-icon.tsx`](../../surfsense_local/frontend/src/features/plugins/plugin-icon.tsx)), name with a Verified checkmark or a Community badge beside it, its description, "By" its author, an `external` entry's note, and **Connect**, or its status once saved. A `token` entry's Connect opens a dialog for the key, labelled in its publisher's words ([`registry-key-dialog.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/registry-key-dialog.tsx)); a `headers` entry's, a field for each header it names, labelled with its `label` or name and its `description` ([`registry-headers-dialog.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/registry-headers-dialog.tsx)). An `oauth` entry's Connect saves it and opens its details in Yours, where the sign-in runs; a `user_client` entry's opens them with **Set up OAuth app**.
-- **Icons.** Each of SurfSense's plugins and each registry plugin shows the icon beside its list; any other shows the first letter of its name ([`plugin-icon.tsx`](../../surfsense_local/frontend/src/features/plugins/plugin-icon.tsx)).
+- **Icons.** Each of SurfSense's plugins and each registry plugin shows the icon beside its list, fading in on its tile once loaded; any other shows the first letter of its name ([`plugin-icon.tsx`](../../surfsense_local/frontend/src/features/plugins/plugin-icon.tsx)).
 - **Add custom plugin** opens one screen, as Claude's own connector dialog does ([`add-plugin-dialog.tsx`](../../surfsense_local/frontend/src/features/plugins/add-custom/add-plugin-dialog.tsx)): Name, Server URL, **Advanced settings** folded, and **Add**. Add puts the server in Yours and the dialog closes on the Yours list, its toast saying how it ended: "{plugin} connected", or "{plugin} added." with what Connect will ask, a sign-in or an OAuth app: one needing nothing, or the key it was given, works at once, and one asking for OAuth waits for **Connect** on its row, as Claude's custom connectors do, which signs in or sets up the user's own OAuth app. Advanced settings holds the connection editor, the same fields as Edit connection ([`connection/`](../../surfsense_local/frontend/src/features/plugins/connection/)): **Headers**, rows of a name and hidden value, up to 10, written as the user would in any MCP client: the first named `Authorization`, its value hinted "Bearer your-key", and renamed for a server reading its own header, such as `x-api-key`. Only rows with a value are sent, and a value needs a name ([`connection-headers.ts`](../../surfsense_local/frontend/src/features/plugins/connection/connection-headers.ts), [`header-rows.tsx`](../../surfsense_local/frontend/src/features/plugins/headers/header-rows.tsx)). The user's own OAuth app is never among them. A refusal shows above them in the app's words by its `code` ([`refusal-message.ts`](../../surfsense_local/frontend/src/features/plugins/refusal-message.ts)), and `plugin_needs_sign_in` opens Advanced settings, since a key fixes it. A provider whose metadata names no registration, as GitHub's, is added as "Needs an OAuth app", and one that refuses SurfSense is found at Connect, which then opens the same three steps. The add form never asks for an OAuth app; a provider SurfSense registers with by itself, Linear's or Notion's, never needs one. A host not allowed raises the egress prompt and retries once allowed. Connecting a key or no-auth plugin from Browse stays on Browse.
 - **A sign-in** started by Connect, Add custom plugin or Sign in opens the provider's page in the browser and shows "Waiting for sign-in…" with Cancel, polling each second until it connects or fails, saying why in the app's words for the API's `code`, in the API's own sentence for a code the app has no words for, and "The sign-in didn’t finish." once the API has forgotten the sign-in ([`use-plugin-sign-in.ts`](../../surfsense_local/frontend/src/features/plugins/sign-in/use-plugin-sign-in.ts), [`sign-in-progress.tsx`](../../surfsense_local/frontend/src/features/plugins/sign-in/sign-in-progress.tsx), [`sign-in-failure-text.ts`](../../surfsense_local/frontend/src/features/plugins/sign-in/sign-in-failure-text.ts)). Every one is followed by Settings › Plugins itself and shown where it started, a row in Yours or a plugin's details, so leaving that row or page does not end it ([`opened-sign-in.ts`](../../surfsense_local/frontend/src/features/plugins/sign-in/opened-sign-in.ts)); Connect in Browse opens the plugin's details in Yours with its sign-in running ([`open-plugin.ts`](../../surfsense_local/frontend/src/features/plugins/open-plugin.ts)); Cancel leaves the plugin saved, "Not connected". Leaving Settings › Plugins abandons it. Electron opens only a URL shaped as SurfSense's authorization request: `response_type=code`, S256 PKCE, SurfSense's `state`, and one of its two callbacks as `redirect_uri` ([`external-url.ts`](../../surfsense_local/electron/src/main/external-url.ts)).
 - **A plugin's public page** opens only from its entry in Browse, in place of the list, with **Back** ([`plugin-public-page.tsx`](../../surfsense_local/frontend/src/features/plugins/public-page/plugin-public-page.tsx)). One page serves first-party and registry plugins, built from the list entry alone, so it is the same whether the plugin is connected or not: its icon and name, beside which a third-party plugin shows a checkmark when it is verified or a small Community badge otherwise, each with a tooltip saying what it means, as on its card ([`publisher-mark.tsx`](../../surfsense_local/frontend/src/features/plugins/public-page/publisher-mark.tsx)), and its publisher, "By SurfSense" or "By" its author, with **Connect**, or its status once saved; Connect on a community plugin first asks, naming who runs it and that the user's requests and sign-in pass through their server ([`community-reminder-dialog.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/community-reminder-dialog.tsx)); **About**, its `about`, a fuller account than its card's `description`, which never stands in for it; **Tools**, its listed tool names as chips in two columns, with no switch; and **Details**: **Publisher**, who runs it, SurfSense or its author; **Authentication** (None, SurfSense license, API key, or "Your {plugin} account" for one that signs in with an account), **Connects to**, its server's host, **Signs in with**, an OAuth plugin's other hosts, **Added**, the month it was added, and its homepage and privacy policy as links that open at once in the browser, by the plugin's id ([egress](egress.md#not-egress)). One Connect serves every listed plugin, on its row and its page ([`connect-listed-plugin.tsx`](../../surfsense_local/frontend/src/features/plugins/browse/connect-listed-plugin.tsx)). A connected plugin's settings are in Yours, below.
diff --git a/surfsense_local/frontend/src/features/plugins/plugin-icon.test.tsx b/surfsense_local/frontend/src/features/plugins/plugin-icon.test.tsx
new file mode 100644
index 0000000000..2fd5bb1eb3
--- /dev/null
+++ b/surfsense_local/frontend/src/features/plugins/plugin-icon.test.tsx
@@ -0,0 +1,47 @@
+import { afterEach, describe, expect, it, vi } from "vitest"
+import { cleanup, fireEvent, screen } from "@testing-library/react"
+
+import { render } from "@/test-utils"
+
+import { PluginIcon } from "./plugin-icon"
+
+function renderIcon() {
+  const { container } = render(
+    <PluginIcon id="notion" name="Notion" hasIcon firstParty />
+  )
+  return container.querySelector("img") as HTMLImageElement
+}
+
+describe("PluginIcon", () => {
+  afterEach(() => {
+    vi.restoreAllMocks()
+    cleanup()
+  })
+
+  it("keeps a logo hidden until it has loaded, then fades it in", () => {
+    const logo = renderIcon()
+    expect(logo.className).toContain("opacity-0")
+
+    fireEvent.load(logo)
+
+    expect(logo.className).toContain("opacity-100")
+    expect(logo.className).toContain("transition-opacity")
+  })
+
+  it("shows a logo the browser already holds at once", () => {
+    vi.spyOn(HTMLImageElement.prototype, "complete", "get").mockReturnValue(
+      true
+    )
+    vi.spyOn(HTMLImageElement.prototype, "naturalWidth", "get").mockReturnValue(
+      32
+    )
+
+    expect(renderIcon().className).toContain("opacity-100")
+  })
+
+  it("falls back to the name's first letter when the logo fails", () => {
+    fireEvent.error(renderIcon())
+
+    expect(screen.getByText("N")).toBeTruthy()
+  })
+})
diff --git a/surfsense_local/frontend/src/features/plugins/plugin-icon.tsx b/surfsense_local/frontend/src/features/plugins/plugin-icon.tsx
index 201f919060..673f289698 100644
--- a/surfsense_local/frontend/src/features/plugins/plugin-icon.tsx
+++ b/surfsense_local/frontend/src/features/plugins/plugin-icon.tsx
@@ -1,4 +1,4 @@
-import { useState } from "react"
+import { useCallback, useState } from "react"
 
 import { cn } from "@/lib/utils"
 
@@ -20,6 +20,11 @@ export function PluginIcon({
   className?: string
 }) {
   const [failed, setFailed] = useState(false)
+  const [loaded, setLoaded] = useState(false)
+  // Set before the first paint, so a logo the browser holds never replays the fade.
+  const showIfAlreadyLoaded = useCallback((logo: HTMLImageElement | null) => {
+    if (logo?.complete && logo.naturalWidth > 0) setLoaded(true)
+  }, [])
   const src = !hasIcon
     ? undefined
     : firstParty
@@ -27,16 +32,25 @@ export function PluginIcon({
       : registryIconUrl(id)
   if (src && !failed) {
     return (
-      <img
-        src={src}
-        alt=""
-        // Registry marks are light or coloured, so they sit on a grey tile, never the bare page.
+      // Registry marks are light or coloured, so they sit on a grey tile, never the bare page.
+      <span
         className={cn(
-          "size-8 shrink-0 rounded-md bg-muted-foreground/20 object-contain p-1 outline -outline-offset-1 outline-black/10 dark:bg-muted dark:outline-white/10",
+          "flex size-8 shrink-0 rounded-md bg-muted-foreground/20 p-1 outline -outline-offset-1 outline-black/10 dark:bg-muted dark:outline-white/10",
           className
         )}
-        onError={() => setFailed(true)}
-      />
+      >
+        <img
+          ref={showIfAlreadyLoaded}
+          src={src}
+          alt=""
+          className={cn(
+            "size-full object-contain transition-opacity duration-150 motion-reduce:transition-none",
+            loaded ? "opacity-100" : "opacity-0"
+          )}
+          onLoad={() => setLoaded(true)}
+          onError={() => setFailed(true)}
+        />
+      </span>
     )
   }
   return (

From 0642cc661c97ca86d0f48a47bb22176e50467385 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 13:41:13 +0530
Subject: [PATCH 07/12] fix(local): refuse a mode pick that lands after the
 model changed

Saving at the pick dropped the check that a mode goes to the model it was picked
on, so a switch mid-save could put it on the next one. The pick now names its
model and is refused if that model is no longer selected. The failure toast also
stops promising the pick lasts until a restart, which a refresh or a refusal breaks.
---
 docs/architecture/model-capabilities.md       |  2 +-
 .../modules/llm/capability/modes/router.py    | 25 +++++--
 .../integration/agent/test_agent_threads.py   |  9 ++-
 .../integration/llm/test_capability_routes.py | 10 ++-
 .../integration/llm/test_pick_chat_mode.py    | 69 +++++++++++++++++--
 .../features/chat/modes/mode-picker.test.tsx  |  7 +-
 .../src/features/chat/modes/save-mode-pick.ts | 17 +++--
 surfsense_local/frontend/translations/de.json |  2 +-
 surfsense_local/frontend/translations/en.json |  2 +-
 surfsense_local/frontend/translations/es.json |  2 +-
 surfsense_local/frontend/translations/fr.json |  2 +-
 surfsense_local/frontend/translations/hi.json |  2 +-
 surfsense_local/frontend/translations/ja.json |  2 +-
 surfsense_local/frontend/translations/ko.json |  2 +-
 .../frontend/translations/pt-BR.json          |  2 +-
 surfsense_local/frontend/translations/ru.json |  2 +-
 .../frontend/translations/zh-CN.json          |  2 +-
 17 files changed, 131 insertions(+), 28 deletions(-)

diff --git a/docs/architecture/model-capabilities.md b/docs/architecture/model-capabilities.md
index 01d78a7290..f7db079982 100644
--- a/docs/architecture/model-capabilities.md
+++ b/docs/architecture/model-capabilities.md
@@ -58,7 +58,7 @@ A chat is Basic (Q&A), the chat engine's retrieval answers with Studio beside it
 
 - `SelectionRead.capability` on `GET` and `PUT /llm/selection/text_gen` (null for other slots): `level`, `label_key` (the ICU select branch), `reason` (`code`, `values`), `note`, `measured` (`key`, `suite`, `assumed`, `suite_version`, `measured_on`, `provider`, `host`, `reads_images`, `passed`, `counted`, `provisional`, or null) and `modes` (`agentic_allowed`, `blocked`, `default_mode`: `basic` or `agentic`, `reason` (`code`, `values`), `remembered_mode`), and `studio`: `full` or `simple`, the Studio path below.
 - `capability_level` on each row of `GET /llm/connections/{id}/models` and `GET /llm/providers/llamacpp/models` that can fill the chat slot.
-- `PUT /llm/selection/text_gen/mode` with `{"mode": "basic" | "agentic"}` saves the pick as the selected model's `remembered_mode` and returns its `modes`; `404` with no chat model chosen, `409` with the gate's `detail.code` for Agentic on a model a gate keeps out ([`modes/router.py`](../../surfsense_local/backend/modules/llm/capability/modes/router.py)).
+- `PUT /llm/selection/text_gen/mode` with `{"mode": "basic" | "agentic", "provider", "connection_id", "name"}` saves the pick as the selected model's `remembered_mode` and returns its `modes`; `404` with no chat model chosen, `409` with `detail.code` `model_changed` when the model named is no longer the selected one, so a pick is never saved onto another model, and `409` with the gate's `detail.code` for Agentic on a model a gate keeps out ([`modes/router.py`](../../surfsense_local/backend/modules/llm/capability/modes/router.py)).
 - `POST /workspaces/{id}/chat/threads` takes `mode` ([chat](chat.md#agent-threads)).
 
 In dev builds only, the model picker tags each chat model by how it did in the Agentic tests: "Agentic", "Agentic, may need nudges", "Low Agentic score" or "Not tested"; production shows no tag. In every build, Settings › Models shows **What this model can do** under the chat model in use ([`model-capability.tsx`](../../surfsense_local/frontend/src/features/models/capability/model-capability.tsx)), closed until opened, so a rating never reads as a verdict beside the model's name. Opened, it says in plain words what Chat offers (both modes and how well, or Basic only and why), what Studio makes (`studio`: full designed pages and Office files, or simple ones and what a stronger model adds), and, last and small, where the rating came from: the cases passed and the day, "From a short test" while the list is provisional, or why it is untested. The row's English note is not shown.
diff --git a/surfsense_local/backend/modules/llm/capability/modes/router.py b/surfsense_local/backend/modules/llm/capability/modes/router.py
index e21875136e..5e3e72a567 100644
--- a/surfsense_local/backend/modules/llm/capability/modes/router.py
+++ b/surfsense_local/backend/modules/llm/capability/modes/router.py
@@ -22,6 +22,10 @@
 
 class ModePick(BaseModel):
     mode: ChatMode
+    # The model it was picked on: the slot may hold another by the time this arrives.
+    provider: str
+    connection_id: int | None
+    name: str
 
 
 @router.put(
@@ -30,20 +34,33 @@ class ModePick(BaseModel):
     summary="Pick the mode new chats on the chosen chat model start in",
 )
 async def pick_mode(payload: ModePick, session: SessionDep) -> ModesRead:
-    return await transact(session, _pick, payload.mode)
+    return await transact(session, _pick, payload)
 
 
-def _pick(session: Session, mode: ChatMode) -> ModesRead:
+def _pick(session: Session, pick: ModePick) -> ModesRead:
     selected = session.get(SelectedModel, ModelType.TEXT_GEN)
     if selected is None:
         raise HTTPException(status.HTTP_404_NOT_FOUND, "no chat model chosen")
+    if (selected.provider, selected.connection_id, selected.name) != (
+        pick.provider,
+        pick.connection_id,
+        pick.name,
+    ):
+        # Never saved: under the selected model it would be a mode nobody chose for it.
+        raise HTTPException(
+            status.HTTP_409_CONFLICT,
+            detail={
+                "code": "model_changed",
+                "message": "The chat model changed before this pick was saved.",
+            },
+        )
     facts = catalog_facts(selected, connection_catalog_provider(session, selected))
     blocked = new_chat_modes(selected, facts).blocked
-    if mode is ChatMode.AGENTIC and blocked is not None:
+    if pick.mode is ChatMode.AGENTIC and blocked is not None:
         # Never saved: the model's default stays a mode it can run.
         raise HTTPException(
             status.HTTP_409_CONFLICT,
             detail={"code": blocked, "message": "This model can't run Agentic."},
         )
-    remember_mode(selected, mode)
+    remember_mode(selected, pick.mode)
     return modes_read(new_chat_modes(selected, facts))
diff --git a/surfsense_local/backend/tests/integration/agent/test_agent_threads.py b/surfsense_local/backend/tests/integration/agent/test_agent_threads.py
index 69d4d701cf..4c3b981300 100644
--- a/surfsense_local/backend/tests/integration/agent/test_agent_threads.py
+++ b/surfsense_local/backend/tests/integration/agent/test_agent_threads.py
@@ -119,8 +119,15 @@ async def test_agentic_asked_for_opens_an_agent_thread_on_an_untested_model(
     """No score blocks Agentic; the user's pick is the model's default from then on."""
     monkeypatch.setattr(get_agent_settings(), "agent_untested_models", False)
 
+    selected = (await agent_api.http.get("/llm/selection/text_gen")).json()
     picked = await agent_api.http.put(
-        "/llm/selection/text_gen/mode", json={"mode": "agentic"}
+        "/llm/selection/text_gen/mode",
+        json={
+            "mode": "agentic",
+            "provider": selected["provider"],
+            "connection_id": selected["connection_id"],
+            "name": selected["name"],
+        },
     )
     picked.raise_for_status()
     chosen = await open_thread(agent_api, mode="agentic")
diff --git a/surfsense_local/backend/tests/integration/llm/test_capability_routes.py b/surfsense_local/backend/tests/integration/llm/test_capability_routes.py
index 7d36663516..6955cac962 100644
--- a/surfsense_local/backend/tests/integration/llm/test_capability_routes.py
+++ b/surfsense_local/backend/tests/integration/llm/test_capability_routes.py
@@ -227,7 +227,15 @@ async def select(name: str) -> dict:
         return chosen.json()["capability"]["modes"]
 
     await select("Qwen3-4B-Q4_K_M")
-    picked = await client.put("/llm/selection/text_gen/mode", json={"mode": "basic"})
+    picked = await client.put(
+        "/llm/selection/text_gen/mode",
+        json={
+            "mode": "basic",
+            "provider": "llamacpp",
+            "connection_id": None,
+            "name": "Qwen3-4B-Q4_K_M",
+        },
+    )
     assert picked.status_code == 200, picked.text
 
     other = await select("Qwen3-1.7B-Q4_K_M")
diff --git a/surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py b/surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py
index bb5f413e97..911de0c712 100644
--- a/surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py
+++ b/surfsense_local/backend/tests/integration/llm/test_pick_chat_mode.py
@@ -6,6 +6,7 @@
 
 from modules.llm.model_type import ModelType
 from modules.llm.models import ProviderConnection, SelectedModel
+from modules.llm.providers import llamacpp
 from shared.config import get_agent_settings
 from shared.db import create_session_factory
 
@@ -24,8 +25,11 @@ def opencode_staged(monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setattr(settings, "agent_untested_models", False)
 
 
-def _select(engine: Engine, name: str, catalog_provider: str, base_url: str) -> None:
-    """A remote selection written as the selection route leaves it; nothing is sent to the host."""
+def _select(engine: Engine, name: str, catalog_provider: str, base_url: str) -> dict:
+    """A remote selection written as the selection route leaves it; nothing is sent to the host.
+
+    Returns the model as a pick names it.
+    """
     with create_session_factory(engine)() as session:
         connection = ProviderConnection(
             label=f"Remote {name}",
@@ -44,6 +48,11 @@ def _select(engine: Engine, name: str, catalog_provider: str, base_url: str) ->
             )
         )
         session.commit()
+        return {
+            "provider": "openai_compatible",
+            "connection_id": connection.id,
+            "name": name,
+        }
 
 
 async def _modes(client: AsyncClient) -> dict:
@@ -56,9 +65,11 @@ async def test_basic_picked_on_an_agentic_model_is_its_default_without_a_chat(
     client: AsyncClient, engine: Engine
 ) -> None:
     """A refresh before the first message keeps the pick."""
-    _select(engine, "anthropic/claude-haiku-4.5", "openrouter", OPENROUTER)
+    model = _select(engine, "anthropic/claude-haiku-4.5", "openrouter", OPENROUTER)
 
-    picked = await client.put("/llm/selection/text_gen/mode", json={"mode": "basic"})
+    picked = await client.put(
+        "/llm/selection/text_gen/mode", json={"mode": "basic", **model}
+    )
 
     assert picked.status_code == 200, picked.text
     modes = await _modes(client)
@@ -69,11 +80,57 @@ async def test_agentic_picked_on_a_model_that_cannot_run_it_is_refused_and_not_k
     client: AsyncClient, engine: Engine
 ) -> None:
     """The catalog says gpt-3.5-turbo makes no tool calls."""
-    _select(engine, "gpt-3.5-turbo", "openai", OPENAI)
+    model = _select(engine, "gpt-3.5-turbo", "openai", OPENAI)
 
-    picked = await client.put("/llm/selection/text_gen/mode", json={"mode": "agentic"})
+    picked = await client.put(
+        "/llm/selection/text_gen/mode", json={"mode": "agentic", **model}
+    )
 
     assert picked.status_code == 409, picked.text
     assert picked.json()["detail"]["code"] == "tool_calls_unsupported"
     modes = await _modes(client)
     assert (modes["default_mode"], modes["remembered_mode"]) == ("basic", None)
+
+
+async def test_pick_made_on_a_model_no_longer_selected_is_refused_and_not_kept(
+    client: AsyncClient, engine: Engine
+) -> None:
+    """The slot took another model, from this device or another, while the pick was on its way."""
+    first = _select(engine, "anthropic/claude-haiku-4.5", "openrouter", OPENROUTER)
+    _select(engine, "anthropic/claude-sonnet-4.5", "openrouter", OPENROUTER)
+
+    picked = await client.put(
+        "/llm/selection/text_gen/mode", json={"mode": "basic", **first}
+    )
+
+    assert picked.status_code == 409, picked.text
+    assert picked.json()["detail"]["code"] == "model_changed"
+    assert (await _modes(client))["remembered_mode"] is None
+
+
+async def test_pick_on_a_local_model_with_no_connection_is_kept(
+    client: AsyncClient, engine: Engine
+) -> None:
+    """A local model has no connection, and the selection read says so with a null."""
+    with create_session_factory(engine)() as session:
+        session.merge(
+            SelectedModel(
+                model_type=ModelType.TEXT_GEN,
+                provider=llamacpp.PROVIDER,
+                name="Qwen3-0.6B",
+            )
+        )
+        session.commit()
+
+    picked = await client.put(
+        "/llm/selection/text_gen/mode",
+        json={
+            "mode": "basic",
+            "provider": llamacpp.PROVIDER,
+            "connection_id": None,
+            "name": "Qwen3-0.6B",
+        },
+    )
+
+    assert picked.status_code == 200, picked.text
+    assert picked.json()["remembered_mode"] == "basic"
diff --git a/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx b/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
index 9ee72ad4d0..992b41d607 100644
--- a/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
+++ b/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
@@ -153,7 +153,12 @@ describe("the composer's mode switch", () => {
         expect.stringContaining("/llm/selection/text_gen/mode"),
         expect.objectContaining({
           method: "PUT",
-          body: JSON.stringify({ mode: "basic" }),
+          body: JSON.stringify({
+            mode: "basic",
+            provider: KIMI.provider,
+            connection_id: KIMI.connection_id,
+            name: KIMI.name,
+          }),
         })
       )
     )
diff --git a/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts b/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
index bc1928179c..706184699d 100644
--- a/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
+++ b/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
@@ -10,11 +10,20 @@ import { requestJson } from "@/lib/api"
 
 import { usePickNewChatMode } from "./new-chat-mode"
 
-function saveModePick(mode: ChatMode): Promise<ChatModes> {
+function saveModePick(
+  model: ModelSelection,
+  mode: ChatMode
+): Promise<ChatModes> {
   return requestJson<ChatModes>("/llm/selection/text_gen/mode", {
     method: "PUT",
     headers: { "Content-Type": "application/json" },
-    body: JSON.stringify({ mode }),
+    // Names the model, so a pick that lands after a switch is refused, not put on the next one.
+    body: JSON.stringify({
+      mode,
+      provider: model.provider,
+      connection_id: model.connection_id,
+      name: model.name,
+    }),
   })
 }
 
@@ -28,7 +37,7 @@ export function useSaveModePick() {
   return useCallback(
     (model: ModelSelection, mode: ChatMode) => {
       pick(model, mode)
-      saveModePick(mode).then(
+      saveModePick(model, mode).then(
         (modes) =>
           client.setQueryData<ModelSelection | null>(
             selectionQueryKey("text_gen"),
@@ -42,7 +51,7 @@ export function useSaveModePick() {
             intl.formatMessage({
               id: "chat_mode_picker_not_saved_toast",
               defaultMessage:
-                "Couldn’t save this mode. New chats use it until you restart.",
+                "Couldn’t save this mode, so it may not stay selected.",
             })
           )
       )
diff --git a/surfsense_local/frontend/translations/de.json b/surfsense_local/frontend/translations/de.json
index c2e7fa5e27..8afd535bb9 100644
--- a/surfsense_local/frontend/translations/de.json
+++ b/surfsense_local/frontend/translations/de.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "Antwortet aus deinen Quellen",
   "chat_mode_picker_change_tooltip": "Basis (Q&A) oder Agentisch",
   "chat_mode_picker_new_chat_label": "Modus für diesen Chat",
-  "chat_mode_picker_not_saved_toast": "Der Modus konnte nicht gespeichert werden. Neue Chats nutzen ihn bis zum Neustart.",
+  "chat_mode_picker_not_saved_toast": "Der Modus konnte nicht gespeichert werden und bleibt möglicherweise nicht ausgewählt.",
   "chat_mode_picker_open_chat_label": "Modus dieses Chats",
   "chat_mode_picker_opens_new_chat_label": "Neuer Chat",
   "chat_mode_picker_thread_aria": "Chatmodus {mode}. Ein Chat behält seinen Modus; starte für einen anderen einen neuen Chat.",
diff --git a/surfsense_local/frontend/translations/en.json b/surfsense_local/frontend/translations/en.json
index 49df3b5b8f..c9f78a24f8 100644
--- a/surfsense_local/frontend/translations/en.json
+++ b/surfsense_local/frontend/translations/en.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "Answers from your sources",
   "chat_mode_picker_change_tooltip": "Basic (Q&A) or Agentic",
   "chat_mode_picker_new_chat_label": "Mode for this chat",
-  "chat_mode_picker_not_saved_toast": "Couldn’t save this mode. New chats use it until you restart.",
+  "chat_mode_picker_not_saved_toast": "Couldn’t save this mode, so it may not stay selected.",
   "chat_mode_picker_open_chat_label": "This chat’s mode",
   "chat_mode_picker_opens_new_chat_label": "New chat",
   "chat_mode_picker_thread_aria": "Chat mode {mode}. A chat keeps its mode; start a new chat for another.",
diff --git a/surfsense_local/frontend/translations/es.json b/surfsense_local/frontend/translations/es.json
index 76b13cdc2e..4506f66b8d 100644
--- a/surfsense_local/frontend/translations/es.json
+++ b/surfsense_local/frontend/translations/es.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "Responde a partir de tus fuentes",
   "chat_mode_picker_change_tooltip": "Básico (P&R) o Agéntico",
   "chat_mode_picker_new_chat_label": "Modo de este chat",
-  "chat_mode_picker_not_saved_toast": "No se pudo guardar este modo. Los chats nuevos lo usarán hasta que reinicies.",
+  "chat_mode_picker_not_saved_toast": "No se pudo guardar este modo, así que puede que no siga seleccionado.",
   "chat_mode_picker_open_chat_label": "Modo de este chat",
   "chat_mode_picker_opens_new_chat_label": "Nuevo chat",
   "chat_mode_picker_thread_aria": "Modo de chat {mode}. Un chat conserva su modo; inicia un chat nuevo para usar otro.",
diff --git a/surfsense_local/frontend/translations/fr.json b/surfsense_local/frontend/translations/fr.json
index d16436156d..d186ab21d3 100644
--- a/surfsense_local/frontend/translations/fr.json
+++ b/surfsense_local/frontend/translations/fr.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "Répond à partir de vos sources",
   "chat_mode_picker_change_tooltip": "Basique (Q&R) ou Agentique",
   "chat_mode_picker_new_chat_label": "Mode de ce chat",
-  "chat_mode_picker_not_saved_toast": "Impossible d’enregistrer ce mode. Les nouveaux chats l’utilisent jusqu’au redémarrage.",
+  "chat_mode_picker_not_saved_toast": "Impossible d’enregistrer ce mode : il risque de ne pas rester sélectionné.",
   "chat_mode_picker_open_chat_label": "Mode de ce chat",
   "chat_mode_picker_opens_new_chat_label": "Nouveau chat",
   "chat_mode_picker_thread_aria": "Mode de chat {mode}. Un chat garde son mode ; démarrez un nouveau chat pour en changer.",
diff --git a/surfsense_local/frontend/translations/hi.json b/surfsense_local/frontend/translations/hi.json
index d06d73ccaf..9a4126f852 100644
--- a/surfsense_local/frontend/translations/hi.json
+++ b/surfsense_local/frontend/translations/hi.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "आपके स्रोतों से जवाब देता है",
   "chat_mode_picker_change_tooltip": "बेसिक (सवाल-जवाब) या एजेंटिक",
   "chat_mode_picker_new_chat_label": "इस चैट का मोड",
-  "chat_mode_picker_not_saved_toast": "यह मोड सहेजा नहीं जा सका। रीस्टार्ट करने तक नई चैट इसी मोड में खुलेंगी।",
+  "chat_mode_picker_not_saved_toast": "यह मोड सहेजा नहीं जा सका, इसलिए हो सकता है कि यह चुना हुआ न रहे।",
   "chat_mode_picker_open_chat_label": "इस चैट का मोड",
   "chat_mode_picker_opens_new_chat_label": "नई चैट",
   "chat_mode_picker_thread_aria": "चैट मोड {mode}. चैट अपना मोड बनाए रखती है; दूसरे मोड के लिए नई चैट शुरू करें।",
diff --git a/surfsense_local/frontend/translations/ja.json b/surfsense_local/frontend/translations/ja.json
index 296a6e0029..1a85a56181 100644
--- a/surfsense_local/frontend/translations/ja.json
+++ b/surfsense_local/frontend/translations/ja.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "ソースをもとに回答",
   "chat_mode_picker_change_tooltip": "ベーシック（Q&A）またはエージェント",
   "chat_mode_picker_new_chat_label": "このチャットのモード",
-  "chat_mode_picker_not_saved_toast": "このモードを保存できませんでした。再起動するまで、新しいチャットはこのモードで開きます。",
+  "chat_mode_picker_not_saved_toast": "このモードを保存できなかったため、選択が維持されない場合があります。",
   "chat_mode_picker_open_chat_label": "このチャットのモード",
   "chat_mode_picker_opens_new_chat_label": "新しいチャット",
   "chat_mode_picker_thread_aria": "チャットモード: {mode}。チャットのモードは変わりません。別のモードにするには新しいチャットを始めてください。",
diff --git a/surfsense_local/frontend/translations/ko.json b/surfsense_local/frontend/translations/ko.json
index 3d14d6c955..96cd221a49 100644
--- a/surfsense_local/frontend/translations/ko.json
+++ b/surfsense_local/frontend/translations/ko.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "소스를 바탕으로 답변",
   "chat_mode_picker_change_tooltip": "기본(Q&A) 또는 에이전트",
   "chat_mode_picker_new_chat_label": "이 채팅의 모드",
-  "chat_mode_picker_not_saved_toast": "이 모드를 저장하지 못했습니다. 다시 시작할 때까지 새 채팅은 이 모드로 열립니다.",
+  "chat_mode_picker_not_saved_toast": "이 모드를 저장하지 못해 선택이 유지되지 않을 수 있습니다.",
   "chat_mode_picker_open_chat_label": "이 채팅의 모드",
   "chat_mode_picker_opens_new_chat_label": "새 채팅",
   "chat_mode_picker_thread_aria": "채팅 모드 {mode}. 채팅은 모드를 유지합니다. 다른 모드를 쓰려면 새 채팅을 시작하세요.",
diff --git a/surfsense_local/frontend/translations/pt-BR.json b/surfsense_local/frontend/translations/pt-BR.json
index 52674766b2..dcf3f85433 100644
--- a/surfsense_local/frontend/translations/pt-BR.json
+++ b/surfsense_local/frontend/translations/pt-BR.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "Responde com base nas suas fontes",
   "chat_mode_picker_change_tooltip": "Básico (P&R) ou Agêntico",
   "chat_mode_picker_new_chat_label": "Modo deste chat",
-  "chat_mode_picker_not_saved_toast": "Não foi possível salvar este modo. Novos chats vão usá-lo até você reiniciar.",
+  "chat_mode_picker_not_saved_toast": "Não foi possível salvar este modo, então ele pode não continuar selecionado.",
   "chat_mode_picker_open_chat_label": "Modo deste chat",
   "chat_mode_picker_opens_new_chat_label": "Novo chat",
   "chat_mode_picker_thread_aria": "Modo do chat {mode}. Um chat mantém o modo; inicie um novo chat para usar outro.",
diff --git a/surfsense_local/frontend/translations/ru.json b/surfsense_local/frontend/translations/ru.json
index b1c221c2fd..e15d75cec4 100644
--- a/surfsense_local/frontend/translations/ru.json
+++ b/surfsense_local/frontend/translations/ru.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "Отвечает по вашим источникам",
   "chat_mode_picker_change_tooltip": "Базовый (вопрос-ответ) или агентный",
   "chat_mode_picker_new_chat_label": "Режим этого чата",
-  "chat_mode_picker_not_saved_toast": "Не удалось сохранить этот режим. Новые чаты будут открываться в нём до перезапуска.",
+  "chat_mode_picker_not_saved_toast": "Не удалось сохранить этот режим, поэтому он может не остаться выбранным.",
   "chat_mode_picker_open_chat_label": "Режим этого чата",
   "chat_mode_picker_opens_new_chat_label": "Новый чат",
   "chat_mode_picker_thread_aria": "Режим чата: {mode}. Чат сохраняет свой режим; для другого начните новый чат.",
diff --git a/surfsense_local/frontend/translations/zh-CN.json b/surfsense_local/frontend/translations/zh-CN.json
index 67a778bbd1..dc586b4fc6 100644
--- a/surfsense_local/frontend/translations/zh-CN.json
+++ b/surfsense_local/frontend/translations/zh-CN.json
@@ -214,7 +214,7 @@
   "chat_mode_basic_short_body": "根据你的来源回答",
   "chat_mode_picker_change_tooltip": "基础（问答）或智能体",
   "chat_mode_picker_new_chat_label": "此对话的模式",
-  "chat_mode_picker_not_saved_toast": "无法保存此模式。重启前，新对话仍会使用它。",
+  "chat_mode_picker_not_saved_toast": "无法保存此模式，因此它可能不会保持选中状态。",
   "chat_mode_picker_open_chat_label": "此对话的模式",
   "chat_mode_picker_opens_new_chat_label": "新对话",
   "chat_mode_picker_thread_aria": "对话模式 {mode}。对话会保持其模式；要换模式，请开始新对话。",

From 350214ce3557a3d9f5662880598b0033932f8afc Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 15:25:39 +0530
Subject: [PATCH 08/12] docs(local): record when a local model's Agentic gate
 is known

Agentic stays selectable on a local model that cannot run it until a chat is
refused. Keeps what was measured about reading it sooner, for when it is built.
---
 docs/architecture/model-capabilities.md | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/docs/architecture/model-capabilities.md b/docs/architecture/model-capabilities.md
index f7db079982..9e06d0b3c1 100644
--- a/docs/architecture/model-capabilities.md
+++ b/docs/architecture/model-capabilities.md
@@ -71,3 +71,5 @@ The composer's switch ([`mode-picker.tsx`](../../surfsense_local/frontend/src/fe
 - No row is measured on this computer, so every local model is `not_measured`; llama.cpp runs need a `served: local` row.
 - The matrix runner writes per-run `result.json` files, not the ladder input; each column is mapped by hand, triage labels included.
 - The sweep's rows come from a committed input, [`capability_list/sweep/openrouter-screen-v1.json`](../../surfsense_local/backend/scripts/capability_list/sweep/openrouter-screen-v1.json): the screen of 7 and 8 Oct 2026 with the 18 re-run models' verdicts in place of their first ones. Its two runs are beside it under `sweep/runs/`, each a `sweep-results.json` and its `attempts.jsonl`; their run folders, about 840 MB of transcripts and files, stay out of the repository. `catalog_rows.py` cannot fold a re-run into a screen, so that input was put together by hand, and no test pins the sweep's rows to it as one pins the ladder's ([#2221](https://github.com/MODSetter/SurfSense/issues/2221)). The 11 models the sweep left unresolved have no row ([#2222](https://github.com/MODSetter/SurfSense/issues/2222)).
+- A local model's tool calls and window are unknown until a chat starts, so Agentic stays selectable on one that cannot run it until its first Agentic chat is refused. A saved Agentic on such a model also survives a refresh or a restart until then. `GET /props?model=<id>&autoload=0` could tell sooner: at llama.cpp b11050 it never loads a model and returned the facts once the model was loaded, adding 5 to 7 ms to a selection read on one 8 GB Mac, with or without replies generating. While the model is not loaded it answers `400`, `500` or `503`, so anything but `200` has to read as unknown. Not built; other hardware is not measured.
+- A server the user connects themselves, such as Ollama or LM Studio, takes the remote path, so the catalog usually says nothing about its models and Agentic stays selectable.

From 5285d9154f8f3a5457eb1e6aa7737173a33d5135 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 16:14:27 +0530
Subject: [PATCH 09/12] feat(local): add a Reduce motion setting
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Settings › General offers Match system, On or Off, kept in the page's
localStorage so the desktop and Docker behave alike. One class on the
document root, set before the first frame, carries the result to
motion-reduce classes, script and a rule that stops other transitions
and animations, leaving spinners, skeletons and scroll fades running.
---
 docs/architecture/overview.md                 |   1 +
 surfsense_local/frontend/index.html           |  11 ++
 .../src/components/typewriter-text.tsx        |   5 +-
 .../src/features/chat/citation-panel.tsx      |   6 +-
 .../motion-preference.test.ts                 | 127 ++++++++++++++++++
 .../motion-preference/motion-preference.ts    |  50 +++++++
 .../welcome-backdrop/gradient-ascii.tsx       |   4 +-
 .../settings/reduce-motion-select.test.tsx    |  58 ++++++++
 .../settings/reduce-motion-select.tsx         |  46 +++++++
 .../src/features/settings/settings-dialog.tsx |  24 ++++
 .../src/features/sources/tree/source-tree.tsx |   6 +-
 .../viewers/flashcards/flashcard-surface.tsx  |   5 +-
 surfsense_local/frontend/src/index.css        |  43 +++---
 surfsense_local/frontend/src/main.tsx         |   2 +
 surfsense_local/frontend/translations/de.json |   5 +
 surfsense_local/frontend/translations/en.json |   5 +
 surfsense_local/frontend/translations/es.json |   5 +
 surfsense_local/frontend/translations/fr.json |   5 +
 surfsense_local/frontend/translations/hi.json |   5 +
 surfsense_local/frontend/translations/ja.json |   5 +
 surfsense_local/frontend/translations/ko.json |   5 +
 .../frontend/translations/pt-BR.json          |   5 +
 surfsense_local/frontend/translations/ru.json |   5 +
 .../frontend/translations/zh-CN.json          |   5 +
 www/content/docs/v2/docker/differences.mdx    |   2 +-
 25 files changed, 413 insertions(+), 27 deletions(-)
 create mode 100644 surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts
 create mode 100644 surfsense_local/frontend/src/features/motion-preference/motion-preference.ts
 create mode 100644 surfsense_local/frontend/src/features/settings/reduce-motion-select.test.tsx
 create mode 100644 surfsense_local/frontend/src/features/settings/reduce-motion-select.tsx

diff --git a/docs/architecture/overview.md b/docs/architecture/overview.md
index da2c9ab41a..152cd08d7e 100644
--- a/docs/architecture/overview.md
+++ b/docs/architecture/overview.md
@@ -111,6 +111,7 @@ The rules that keep the processes out of each other's way:
 - TanStack Query holds most server state: threads and messages, the citation panel, the model catalog and settings, the Studio formats and artifact list, and an open artifact with what its viewer reads. An `artifacts` event invalidates the list and every open artifact under one key prefix ([`features/studio/query-keys.ts`](../../surfsense_local/frontend/src/features/studio/query-keys.ts)), so a regenerated artifact that is open shows its new generation. The sources list is component state, reloaded on the events above.
 - There is no router. [`app/app-bootstrap.tsx`](../../surfsense_local/frontend/src/app/app-bootstrap.tsx) asks `GET /llm/onboarding`, then renders either onboarding ([`local-models/selection.md`](local-models/selection.md)) or the lazily loaded dashboard: the workspace rail, threads and sources on the left, the conversation in the middle, and Studio, artifacts and the citation panel on the right. The left and right columns resize by dragging their inner edge, and the conversation keeps at least 520 px ([studio](studio.md#the-studio-panel)).
 - While startup waits on the API, [`features/startup/`](../../surfsense_local/frontend/src/features/startup/) draws the dashboard's constant parts, with nothing to press: the columns at their saved widths, the "SurfSense" and "Studio" headings, and either a new chat's box or an open thread's message skeleton, as the last view was; a thread's composer arrives with its messages. The new chat's box takes typing but sends nothing, and the text goes to the dashboard's composer through a store `AppBootstrap` owns and provides to both (`startup-input`). The composer sets the text with `composer.setText`, then, once it shows, places the caret and tells the `@` trigger where it is through assistant-ui's trigger context (`resource.setCursorPosition`): the list tracks its own cursor, apart from the box's. Startup waits for an input method's composition to end first. One `@` rule (`mention-matcher.ts`), given to assistant-ui as its matcher, decides in both places whether the caret is in a mention. When new chats open in Agentic, which the dashboard stores for the next launch, `@` in the box shows a skeleton of the plugin list, which startup starts loading early. The dashboard draws the same parts with the same components (`composer-surface.tsx`, `conversation-scroll.tsx`, `mention-list.tsx`, rendered around assistant-ui's parts with Base UI's `useRender`), and both text boxes grow with the same autosizing textarea (`react-textarea-autosize`, which assistant-ui's input is built on), so nothing moves when it takes over; its contents fade in over 200 ms, or appear at once if the wait was under 150 ms or motion is reduced. The frame is drawn only when onboarding was last seen done on the same database, which Electron names by the database file's identity (`window.surfsense.databaseIdentity()`), so a database made since, a first launch and onboarding get the plain background. Docker has no bridge and always gets the plain background. A test walks the frame's imports with TypeScript's parser and keeps them off the chat runtime and every feature's `api.ts`, so it stays in the first chunk. No test measures the two frames' layout in a browser yet.
+- Settings › General's Reduce motion is Match system, On or Off, kept in the page's `localStorage`, so the desktop and Docker behave alike. [`features/motion-preference/`](../../surfsense_local/frontend/src/features/motion-preference/) turns it and the OS setting into one `reduce-motion` class on the document root, and `index.html` sets the class before the first frame. `motion-reduce:` and `motion-safe:` read that class, script reads `motionIsReduced()`, and one rule in `index.css` stops every other transition and animation except spinners, skeletons and scroll fades. Off can't reach the OS-only rules that sonner and shadcn's shimmer ship, so with the OS asking to reduce motion, toasts and the shimmer stay still.
 - Code is grouped by feature under `src/features/`. Each HTTP contract lives in one feature's `api.ts`, next to its hooks and components, and other features import it from there.
 
 ## Codemap
diff --git a/surfsense_local/frontend/index.html b/surfsense_local/frontend/index.html
index c6c64af3bc..08367575d1 100644
--- a/surfsense_local/frontend/index.html
+++ b/surfsense_local/frontend/index.html
@@ -30,6 +30,17 @@
         } catch {
           // no-op: fall back to the CSS defaults if storage/matchMedia throws
         }
+        // Same rule as features/motion-preference, for the first frame.
+        try {
+          const motion = localStorage.getItem("surfsense:reduce-motion:v1")
+          const reduce =
+            motion === "on" ||
+            (motion !== "off" &&
+              window.matchMedia("(prefers-reduced-motion: reduce)").matches)
+          if (reduce) document.documentElement.classList.add("reduce-motion")
+        } catch {
+          // no-op: the bundle sets it again once it loads
+        }
       })()
     </script>
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
diff --git a/surfsense_local/frontend/src/components/typewriter-text.tsx b/surfsense_local/frontend/src/components/typewriter-text.tsx
index 3dfc49d09d..f3b7a0a03d 100644
--- a/surfsense_local/frontend/src/components/typewriter-text.tsx
+++ b/surfsense_local/frontend/src/components/typewriter-text.tsx
@@ -1,6 +1,6 @@
 import { useEffect, useState } from "react"
 
-const REDUCED_MOTION_QUERY = "(prefers-reduced-motion: reduce)"
+import { motionIsReduced } from "@/features/motion-preference/motion-preference"
 
 function AnimatedText({
   text,
@@ -44,8 +44,7 @@ export function TypewriterText({
   animate?: boolean
   onComplete?: () => void
 }) {
-  const reducedMotion =
-    window.matchMedia?.(REDUCED_MOTION_QUERY).matches ?? false
+  const reducedMotion = motionIsReduced()
   return (
     <>
       <span className="sr-only">{text}</span>
diff --git a/surfsense_local/frontend/src/features/chat/citation-panel.tsx b/surfsense_local/frontend/src/features/chat/citation-panel.tsx
index 5904229acb..68e703bac5 100644
--- a/surfsense_local/frontend/src/features/chat/citation-panel.tsx
+++ b/surfsense_local/frontend/src/features/chat/citation-panel.tsx
@@ -4,6 +4,7 @@ import { useEffect, useMemo, useRef } from "react"
 import { Button } from "@/components/ui/button"
 import { DetailPanel } from "@/components/ui/detail-panel"
 import { Spinner } from "@/components/ui/spinner"
+import { motionIsReduced } from "@/features/motion-preference/motion-preference"
 import { getDocumentByChunk } from "@/features/sources/api"
 import { intl } from "@/i18n/intl"
 
@@ -47,7 +48,10 @@ export function CitationPanel({
         target.getBoundingClientRect().top -
         container.getBoundingClientRect().top +
         container.scrollTop
-      container.scrollTo({ top: Math.max(0, offset - 16), behavior: "smooth" })
+      container.scrollTo({
+        top: Math.max(0, offset - 16),
+        behavior: motionIsReduced() ? "auto" : "smooth",
+      })
     })
     return () => cancelAnimationFrame(id)
   }, [cited])
diff --git a/surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts b/surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts
new file mode 100644
index 0000000000..7215b732ab
--- /dev/null
+++ b/surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts
@@ -0,0 +1,127 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest"
+
+import indexHtml from "../../../index.html?raw"
+
+import {
+  followMotionPreference,
+  MOTION_PREFERENCE_KEY,
+  motionIsReduced,
+  setMotionPreference,
+} from "./motion-preference"
+
+const CLASS = "reduce-motion"
+const root = document.documentElement
+
+function osAsksToReduce(matches: boolean) {
+  const listeners = new Set<() => void>()
+  const query = {
+    matches,
+    addEventListener: (_: string, listener: () => void) =>
+      listeners.add(listener),
+    removeEventListener: (_: string, listener: () => void) =>
+      listeners.delete(listener),
+  }
+  vi.stubGlobal("matchMedia", (text: string) =>
+    text.includes("reduced-motion") ? query : { matches: false }
+  )
+  return {
+    change(next: boolean) {
+      query.matches = next
+      listeners.forEach((listener) => listener())
+    },
+  }
+}
+
+// The cases every reader must agree on: [saved choice, OS asks, reduced].
+const CASES: [string | null, boolean, boolean][] = [
+  [null, false, false],
+  [null, true, true],
+  ["system", false, false],
+  ["system", true, true],
+  ["on", false, true],
+  ["on", true, true],
+  ["off", false, false],
+  ["off", true, false],
+  ["nonsense", true, true],
+  ["nonsense", false, false],
+]
+
+function save(value: string | null) {
+  if (value !== null) localStorage.setItem(MOTION_PREFERENCE_KEY, value)
+}
+
+beforeEach(() => {
+  localStorage.clear()
+  root.classList.remove(CLASS)
+})
+
+afterEach(() => {
+  vi.unstubAllGlobals()
+  root.classList.remove(CLASS, "light", "dark")
+})
+
+describe("motion preference", () => {
+  it.each(CASES)(
+    "with %s saved and the OS asking %s, reduced is %s",
+    (saved, asks, reduced) => {
+      save(saved)
+      osAsksToReduce(asks)
+
+      expect(motionIsReduced()).toBe(reduced)
+      followMotionPreference()
+      expect(root.classList.contains(CLASS)).toBe(reduced)
+    }
+  )
+
+  it.each(CASES)(
+    "index.html's first-frame script agrees with %s saved and the OS asking %s",
+    (saved, asks, reduced) => {
+      save(saved)
+      osAsksToReduce(asks)
+
+      const script = indexHtml.match(/<script>([\s\S]*?)<\/script>/)?.[1] ?? ""
+      new Function(script)()
+
+      expect(root.classList.contains(CLASS)).toBe(reduced)
+    }
+  )
+
+  it("takes effect at once when the choice changes, and survives a reload", () => {
+    osAsksToReduce(false)
+
+    setMotionPreference("on")
+    expect(root.classList.contains(CLASS)).toBe(true)
+    expect(motionIsReduced()).toBe(true)
+
+    setMotionPreference("off")
+    expect(root.classList.contains(CLASS)).toBe(false)
+    expect(localStorage.getItem(MOTION_PREFERENCE_KEY)).toBe("off")
+  })
+
+  it("follows the OS while it matches the system, and ignores it otherwise", () => {
+    const os = osAsksToReduce(false)
+    followMotionPreference()
+
+    os.change(true)
+    expect(root.classList.contains(CLASS)).toBe(true)
+
+    setMotionPreference("off")
+    os.change(true)
+    expect(root.classList.contains(CLASS)).toBe(false)
+  })
+
+  it("follows a choice made in another tab", () => {
+    osAsksToReduce(false)
+    followMotionPreference()
+
+    localStorage.setItem(MOTION_PREFERENCE_KEY, "on")
+    window.dispatchEvent(
+      new StorageEvent("storage", {
+        key: MOTION_PREFERENCE_KEY,
+        storageArea: localStorage,
+      })
+    )
+
+    expect(root.classList.contains(CLASS)).toBe(true)
+  })
+})
diff --git a/surfsense_local/frontend/src/features/motion-preference/motion-preference.ts b/surfsense_local/frontend/src/features/motion-preference/motion-preference.ts
new file mode 100644
index 0000000000..f6dd1d7d87
--- /dev/null
+++ b/surfsense_local/frontend/src/features/motion-preference/motion-preference.ts
@@ -0,0 +1,50 @@
+export type MotionPreference = "system" | "on" | "off"
+
+// index.html reads these before the bundle loads, so the first frame is right;
+// its test fails if the two disagree.
+export const MOTION_PREFERENCE_KEY = "surfsense:reduce-motion:v1"
+const REDUCED_MOTION_QUERY = "(prefers-reduced-motion: reduce)"
+const ROOT_CLASS = "reduce-motion"
+
+export function readMotionPreference(): MotionPreference {
+  try {
+    const saved = localStorage.getItem(MOTION_PREFERENCE_KEY)
+    return saved === "on" || saved === "off" ? saved : "system"
+  } catch {
+    return "system"
+  }
+}
+
+function systemAsksToReduce(): boolean {
+  return window.matchMedia?.(REDUCED_MOTION_QUERY).matches ?? false
+}
+
+/** For code that animates from script; CSS reads the root class instead. */
+export function motionIsReduced(): boolean {
+  const preference = readMotionPreference()
+  return preference === "system" ? systemAsksToReduce() : preference === "on"
+}
+
+function applyToRoot(): void {
+  document.documentElement.classList.toggle(ROOT_CLASS, motionIsReduced())
+}
+
+export function setMotionPreference(preference: MotionPreference): void {
+  try {
+    localStorage.setItem(MOTION_PREFERENCE_KEY, preference)
+  } catch {
+    // Without storage, the choice cannot be kept.
+  }
+  applyToRoot()
+}
+
+/** Keeps the root class right when the OS setting or another tab's choice changes. */
+export function followMotionPreference(): void {
+  applyToRoot()
+  window
+    .matchMedia?.(REDUCED_MOTION_QUERY)
+    .addEventListener?.("change", applyToRoot)
+  window.addEventListener("storage", (event) => {
+    if (event.key === MOTION_PREFERENCE_KEY || event.key === null) applyToRoot()
+  })
+}
diff --git a/surfsense_local/frontend/src/features/onboarding/welcome-backdrop/gradient-ascii.tsx b/surfsense_local/frontend/src/features/onboarding/welcome-backdrop/gradient-ascii.tsx
index d285cdc873..bbce5431ed 100644
--- a/surfsense_local/frontend/src/features/onboarding/welcome-backdrop/gradient-ascii.tsx
+++ b/surfsense_local/frontend/src/features/onboarding/welcome-backdrop/gradient-ascii.tsx
@@ -1,5 +1,6 @@
 import { useEffect, useRef } from "react"
 
+import { motionIsReduced } from "@/features/motion-preference/motion-preference"
 import { cn } from "@/lib/utils"
 
 import {
@@ -163,8 +164,7 @@ export function GradientAscii({
       return rgb
     }
 
-    const reduce =
-      window.matchMedia?.("(prefers-reduced-motion: reduce)").matches ?? false
+    const reduce = motionIsReduced()
     const view = { width: 1, height: 1, ratio: 1 }
     const ramp = new Float32Array(MAX_COLORS * 3)
     const clickData = new Float32Array(MAX_CLICKS * 4)
diff --git a/surfsense_local/frontend/src/features/settings/reduce-motion-select.test.tsx b/surfsense_local/frontend/src/features/settings/reduce-motion-select.test.tsx
new file mode 100644
index 0000000000..4bfe527953
--- /dev/null
+++ b/surfsense_local/frontend/src/features/settings/reduce-motion-select.test.tsx
@@ -0,0 +1,58 @@
+import { cleanup, render, screen } from "@testing-library/react"
+import userEvent from "@testing-library/user-event"
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest"
+
+import { MOTION_PREFERENCE_KEY } from "@/features/motion-preference/motion-preference"
+
+import { ReduceMotionSelect } from "./reduce-motion-select"
+
+beforeEach(() => {
+  localStorage.clear()
+  vi.stubGlobal("matchMedia", () => ({ matches: false }))
+})
+
+afterEach(() => {
+  cleanup()
+  vi.unstubAllGlobals()
+  document.documentElement.classList.remove("reduce-motion")
+})
+
+describe("ReduceMotionSelect", () => {
+  it("offers Match system, On and Off, and starts on Match system", () => {
+    render(<ReduceMotionSelect aria-label="Reduce motion" />)
+
+    const select = screen.getByRole<HTMLSelectElement>("combobox", {
+      name: "Reduce motion",
+    })
+    expect(select.value).toBe("system")
+    expect(
+      screen.getAllByRole("option").map((option) => option.textContent)
+    ).toEqual(["Match system", "On", "Off"])
+  })
+
+  it("shows the saved choice", () => {
+    localStorage.setItem(MOTION_PREFERENCE_KEY, "off")
+    render(<ReduceMotionSelect aria-label="Reduce motion" />)
+
+    expect(
+      screen.getByRole<HTMLSelectElement>("combobox", {
+        name: "Reduce motion",
+      }).value
+    ).toBe("off")
+  })
+
+  it("reduces motion as soon as On is picked", async () => {
+    const user = userEvent.setup()
+    render(<ReduceMotionSelect aria-label="Reduce motion" />)
+
+    await user.selectOptions(
+      screen.getByRole("combobox", { name: "Reduce motion" }),
+      "on"
+    )
+
+    expect(document.documentElement.classList.contains("reduce-motion")).toBe(
+      true
+    )
+    expect(localStorage.getItem(MOTION_PREFERENCE_KEY)).toBe("on")
+  })
+})
diff --git a/surfsense_local/frontend/src/features/settings/reduce-motion-select.tsx b/surfsense_local/frontend/src/features/settings/reduce-motion-select.tsx
new file mode 100644
index 0000000000..2c296b586e
--- /dev/null
+++ b/surfsense_local/frontend/src/features/settings/reduce-motion-select.tsx
@@ -0,0 +1,46 @@
+import { useState, type ComponentProps } from "react"
+
+import { Select } from "@/components/ui/select"
+import {
+  readMotionPreference,
+  setMotionPreference,
+  type MotionPreference,
+} from "@/features/motion-preference/motion-preference"
+import { intl } from "@/i18n/intl"
+
+export function ReduceMotionSelect(
+  props: Omit<ComponentProps<typeof Select>, "value" | "onChange">
+) {
+  const [preference, setPreference] = useState(readMotionPreference)
+
+  return (
+    <Select
+      {...props}
+      value={preference}
+      onChange={(event) => {
+        const next = event.target.value as MotionPreference
+        setPreference(next)
+        setMotionPreference(next)
+      }}
+    >
+      <option value="system">
+        {intl.formatMessage({
+          id: "settings_general_reduce_motion_system_label",
+          defaultMessage: "Match system",
+        })}
+      </option>
+      <option value="on">
+        {intl.formatMessage({
+          id: "settings_general_reduce_motion_on_label",
+          defaultMessage: "On",
+        })}
+      </option>
+      <option value="off">
+        {intl.formatMessage({
+          id: "settings_general_reduce_motion_off_label",
+          defaultMessage: "Off",
+        })}
+      </option>
+    </Select>
+  )
+}
diff --git a/surfsense_local/frontend/src/features/settings/settings-dialog.tsx b/surfsense_local/frontend/src/features/settings/settings-dialog.tsx
index 2977acebbd..07467c8e45 100644
--- a/surfsense_local/frontend/src/features/settings/settings-dialog.tsx
+++ b/surfsense_local/frontend/src/features/settings/settings-dialog.tsx
@@ -52,6 +52,7 @@ import { cn } from "@/lib/utils"
 
 import { AppearanceToggle } from "./appearance-toggle"
 import { LanguageSelect } from "./language-select"
+import { ReduceMotionSelect } from "./reduce-motion-select"
 import { AudioModelsSettings } from "./models/audio-models-settings"
 import { ChatModelsSettings } from "./models/chat-models-settings"
 import { ImageEditModelsSettings } from "./models/image-edit-models-settings"
@@ -139,6 +140,29 @@ function GeneralSettings({
           className="w-40"
         />
       </div>
+      <div className="mt-8 flex items-center justify-between gap-8">
+        <div className="flex flex-col gap-1">
+          <h3
+            id="settings-general-reduce-motion"
+            className="text-sm font-medium"
+          >
+            {intl.formatMessage({
+              id: "settings_general_reduce_motion_title",
+              defaultMessage: "Reduce motion",
+            })}
+          </h3>
+          <p className="text-sm text-pretty text-muted-foreground">
+            {intl.formatMessage({
+              id: "settings_general_reduce_motion_body",
+              defaultMessage: "Keep animations to a minimum on this device.",
+            })}
+          </p>
+        </div>
+        <ReduceMotionSelect
+          aria-labelledby="settings-general-reduce-motion"
+          className="w-40"
+        />
+      </div>
       <div className="mt-8 flex items-center justify-between gap-8">
         <div className="flex flex-col gap-1">
           <div className="flex items-center gap-1">
diff --git a/surfsense_local/frontend/src/features/sources/tree/source-tree.tsx b/surfsense_local/frontend/src/features/sources/tree/source-tree.tsx
index d0ea9c25e9..bf01fcada6 100644
--- a/surfsense_local/frontend/src/features/sources/tree/source-tree.tsx
+++ b/surfsense_local/frontend/src/features/sources/tree/source-tree.tsx
@@ -9,6 +9,7 @@ import {
 } from "react"
 
 import { getFileViewer } from "@/features/file-viewers/registry"
+import { motionIsReduced } from "@/features/motion-preference/motion-preference"
 import { useStableCallback } from "@/hooks/use-stable-callback"
 import { intl } from "@/i18n/intl"
 
@@ -100,7 +101,10 @@ export const SourceTree = memo(function SourceTree({
     if (highlightedDocumentId === null) return
     rowElements.current
       .get(documentKey(highlightedDocumentId))
-      ?.scrollIntoView({ behavior: "smooth", block: "nearest" })
+      ?.scrollIntoView({
+        behavior: motionIsReduced() ? "auto" : "smooth",
+        block: "nearest",
+      })
   }, [highlightedDocumentId])
 
   const focusRow = (row: TreeRow | undefined) => {
diff --git a/surfsense_local/frontend/src/features/studio/viewers/flashcards/flashcard-surface.tsx b/surfsense_local/frontend/src/features/studio/viewers/flashcards/flashcard-surface.tsx
index 05865a61e9..36f448150c 100644
--- a/surfsense_local/frontend/src/features/studio/viewers/flashcards/flashcard-surface.tsx
+++ b/surfsense_local/frontend/src/features/studio/viewers/flashcards/flashcard-surface.tsx
@@ -1,13 +1,13 @@
 import type { ReactNode } from "react"
 
 import { Button } from "@/components/ui/button"
+import { motionIsReduced } from "@/features/motion-preference/motion-preference"
 import { intl } from "@/i18n/intl"
 import { cn } from "@/lib/utils"
 
 // Ported from surfsense_web's flip card, minus the `motion/react` dependency
 // this app doesn't have — a plain CSS transform transition gives the same
 // 3D flip without pulling in framer-motion for one component.
-const REDUCED_MOTION_QUERY = "(prefers-reduced-motion: reduce)"
 
 function Face({
   children,
@@ -44,8 +44,7 @@ export function FlashcardSurface({
   revealed: boolean
   onFlip: () => void
 }) {
-  const reducedMotion =
-    window.matchMedia?.(REDUCED_MOTION_QUERY).matches ?? false
+  const reducedMotion = motionIsReduced()
 
   if (reducedMotion) {
     return (
diff --git a/surfsense_local/frontend/src/index.css b/surfsense_local/frontend/src/index.css
index 55022e73e8..0bea125bdf 100644
--- a/surfsense_local/frontend/src/index.css
+++ b/surfsense_local/frontend/src/index.css
@@ -11,6 +11,9 @@
 @source "../node_modules/@streamdown/math/dist/*.js";
 
 @custom-variant dark (&:is(.dark *));
+/* Settings › General decides, not only the OS (`features/motion-preference`). */
+@custom-variant motion-reduce (&:is(.reduce-motion *));
+@custom-variant motion-safe (&:not(.reduce-motion *));
 
 @theme inline {
   --font-heading: var(--font-sans);
@@ -433,12 +436,6 @@
   animation: ss-reveal-in 200ms cubic-bezier(0.215, 0.61, 0.355, 1);
 }
 
-@media (prefers-reduced-motion: reduce) {
-  .ss-reveal > :not([data-reveal-steady]) {
-    animation: none;
-  }
-}
-
 /* The thinking indicator (`features/chat/thinking-indicator.tsx`), ported
  * from surfsense_web. Dark keeps the web's light-on-screen glow; light draws
  * the same dots in ink, since screen blending on a pale ground erases them. */
@@ -701,16 +698,30 @@
   }
 }
 
-@media (prefers-reduced-motion: reduce) {
-  .ss-thinking-dot,
-  .ss-thinking-orbit,
-  .ss-thinking-pulse,
-  .ss-thinking-breathe-primary,
-  .ss-thinking-breathe-secondary,
-  .ss-thinking-light,
-  .ss-thinking-goo {
-    animation: none;
-  }
+/* Reduced motion, for everything not written with it in mind. Spinners and
+ * skeletons stay, since they say work is under way; scroll-fade is a mask
+ * driven by the scroll position, not motion. */
+.reduce-motion *,
+.reduce-motion *::before,
+.reduce-motion *::after {
+  transition: none !important;
+}
+
+.reduce-motion
+  *:not(
+    [class*="animate-spin"],
+    [class*="animate-pulse"],
+    [class*="scroll-fade"]
+  ),
+.reduce-motion *::before,
+.reduce-motion *::after {
+  animation: none !important;
+}
+
+/* shadcn's own reduced-motion rule for it reads only the OS. */
+.reduce-motion .shimmer {
+  background-image: none;
+  -webkit-text-fill-color: currentColor;
 }
 
 /* Firefox before 159 has no scroll-driven animations, so shadcn's scroll-fade
diff --git a/surfsense_local/frontend/src/main.tsx b/surfsense_local/frontend/src/main.tsx
index a4e8e35254..98facfcfd0 100644
--- a/surfsense_local/frontend/src/main.tsx
+++ b/surfsense_local/frontend/src/main.tsx
@@ -13,6 +13,7 @@ import { TooltipProvider } from "@/components/ui/tooltip.tsx"
 import { EgressPrompt } from "@/features/egress/egress-prompt.tsx"
 import { IssueReportDialog } from "@/features/feedback/issue-report-dialog.tsx"
 import { SidecarCrashReporter } from "@/features/feedback/sidecar-crash-reporter.tsx"
+import { followMotionPreference } from "@/features/motion-preference/motion-preference.ts"
 import { InstallFeed } from "@/features/models/local/installs/install-feed.tsx"
 import { MenuUpdateCheck } from "@/features/updates/menu-update-check.tsx"
 import { intl } from "@/i18n/intl.ts"
@@ -32,6 +33,7 @@ if (window.surfsense?.platform) {
 }
 
 followLocale()
+followMotionPreference()
 guardFileDrops(window)
 
 // Open over any dialog as its nested dialog, or on their own when none is open.
diff --git a/surfsense_local/frontend/translations/de.json b/surfsense_local/frontend/translations/de.json
index 8afd535bb9..1f05129300 100644
--- a/surfsense_local/frontend/translations/de.json
+++ b/surfsense_local/frontend/translations/de.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "Wähle die Sprache, die SurfSense auf diesem Gerät verwendet.",
   "settings_general_language_system_label": "Wie System",
   "settings_general_language_title": "Sprache",
+  "settings_general_reduce_motion_body": "Halte Animationen auf diesem Gerät auf ein Minimum.",
+  "settings_general_reduce_motion_off_label": "Aus",
+  "settings_general_reduce_motion_on_label": "An",
+  "settings_general_reduce_motion_system_label": "Wie System",
+  "settings_general_reduce_motion_title": "Bewegung reduzieren",
   "settings_general_title": "Allgemein",
   "settings_image_edit_models_body": "Das Modell, das Bilder bearbeitet. Führe eines auf diesem Computer aus oder nutze eines von einem Server.",
   "settings_image_edit_models_title": "Modelle für Bildbearbeitung",
diff --git a/surfsense_local/frontend/translations/en.json b/surfsense_local/frontend/translations/en.json
index c9f78a24f8..03c3c8adb6 100644
--- a/surfsense_local/frontend/translations/en.json
+++ b/surfsense_local/frontend/translations/en.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "Choose the language SurfSense uses on this device.",
   "settings_general_language_system_label": "Match system",
   "settings_general_language_title": "Language",
+  "settings_general_reduce_motion_body": "Keep animations to a minimum on this device.",
+  "settings_general_reduce_motion_off_label": "Off",
+  "settings_general_reduce_motion_on_label": "On",
+  "settings_general_reduce_motion_system_label": "Match system",
+  "settings_general_reduce_motion_title": "Reduce motion",
   "settings_general_title": "General",
   "settings_image_edit_models_body": "The model that edits images. Run one on this computer, or use one from a server.",
   "settings_image_edit_models_title": "Image editing models",
diff --git a/surfsense_local/frontend/translations/es.json b/surfsense_local/frontend/translations/es.json
index 4506f66b8d..477cc1097a 100644
--- a/surfsense_local/frontend/translations/es.json
+++ b/surfsense_local/frontend/translations/es.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "Elige el idioma que SurfSense usa en este dispositivo.",
   "settings_general_language_system_label": "Igual que el sistema",
   "settings_general_language_title": "Idioma",
+  "settings_general_reduce_motion_body": "Reduce las animaciones al mínimo en este dispositivo.",
+  "settings_general_reduce_motion_off_label": "Desactivado",
+  "settings_general_reduce_motion_on_label": "Activado",
+  "settings_general_reduce_motion_system_label": "Igual que el sistema",
+  "settings_general_reduce_motion_title": "Reducir movimiento",
   "settings_general_title": "General",
   "settings_image_edit_models_body": "El modelo que edita imágenes. Ejecuta uno en este equipo o usa uno de un servidor.",
   "settings_image_edit_models_title": "Modelos de edición de imágenes",
diff --git a/surfsense_local/frontend/translations/fr.json b/surfsense_local/frontend/translations/fr.json
index d186ab21d3..e6580e22e0 100644
--- a/surfsense_local/frontend/translations/fr.json
+++ b/surfsense_local/frontend/translations/fr.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "Choisissez la langue utilisée par SurfSense sur cet appareil.",
   "settings_general_language_system_label": "Comme le système",
   "settings_general_language_title": "Langue",
+  "settings_general_reduce_motion_body": "Limitez les animations au minimum sur cet appareil.",
+  "settings_general_reduce_motion_off_label": "Désactivé",
+  "settings_general_reduce_motion_on_label": "Activé",
+  "settings_general_reduce_motion_system_label": "Comme le système",
+  "settings_general_reduce_motion_title": "Réduire les animations",
   "settings_general_title": "Général",
   "settings_image_edit_models_body": "Le modèle qui modifie les images. Exécutez-en un sur cet ordinateur ou utilisez-en un depuis un serveur.",
   "settings_image_edit_models_title": "Modèles de modification d’images",
diff --git a/surfsense_local/frontend/translations/hi.json b/surfsense_local/frontend/translations/hi.json
index 9a4126f852..19ae883ad7 100644
--- a/surfsense_local/frontend/translations/hi.json
+++ b/surfsense_local/frontend/translations/hi.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "चुनें कि इस डिवाइस पर SurfSense कौन सी भाषा इस्तेमाल करे।",
   "settings_general_language_system_label": "सिस्टम के अनुसार",
   "settings_general_language_title": "भाषा",
+  "settings_general_reduce_motion_body": "इस डिवाइस पर ऐनिमेशन कम से कम रखें।",
+  "settings_general_reduce_motion_off_label": "बंद",
+  "settings_general_reduce_motion_on_label": "चालू",
+  "settings_general_reduce_motion_system_label": "सिस्टम के अनुसार",
+  "settings_general_reduce_motion_title": "मोशन कम करें",
   "settings_general_title": "सामान्य",
   "settings_image_edit_models_body": "वह मॉडल जो इमेज एडिट करता है। इसे इस कंप्यूटर पर चलाएं, या किसी सर्वर से इस्तेमाल करें।",
   "settings_image_edit_models_title": "इमेज एडिटिंग मॉडल",
diff --git a/surfsense_local/frontend/translations/ja.json b/surfsense_local/frontend/translations/ja.json
index 1a85a56181..f4791332c0 100644
--- a/surfsense_local/frontend/translations/ja.json
+++ b/surfsense_local/frontend/translations/ja.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "このデバイスでSurfSenseが使う言語を選びます。",
   "settings_general_language_system_label": "システムに合わせる",
   "settings_general_language_title": "言語",
+  "settings_general_reduce_motion_body": "このデバイスでのアニメーションを最小限にします。",
+  "settings_general_reduce_motion_off_label": "オフ",
+  "settings_general_reduce_motion_on_label": "オン",
+  "settings_general_reduce_motion_system_label": "システムに合わせる",
+  "settings_general_reduce_motion_title": "動きを減らす",
   "settings_general_title": "一般",
   "settings_image_edit_models_body": "画像を編集するモデルです。このコンピューターで実行するか、サーバー上のモデルを使います。",
   "settings_image_edit_models_title": "画像編集モデル",
diff --git a/surfsense_local/frontend/translations/ko.json b/surfsense_local/frontend/translations/ko.json
index 96cd221a49..0683cb9c8d 100644
--- a/surfsense_local/frontend/translations/ko.json
+++ b/surfsense_local/frontend/translations/ko.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "이 기기에서 SurfSense가 사용할 언어를 선택합니다.",
   "settings_general_language_system_label": "시스템에 맞춤",
   "settings_general_language_title": "언어",
+  "settings_general_reduce_motion_body": "이 기기에서 애니메이션을 최소화합니다.",
+  "settings_general_reduce_motion_off_label": "꺼짐",
+  "settings_general_reduce_motion_on_label": "켜짐",
+  "settings_general_reduce_motion_system_label": "시스템에 맞춤",
+  "settings_general_reduce_motion_title": "동작 줄이기",
   "settings_general_title": "일반",
   "settings_image_edit_models_body": "이미지를 편집하는 모델입니다. 이 컴퓨터에서 실행하거나, 서버의 모델을 사용하세요.",
   "settings_image_edit_models_title": "이미지 편집 모델",
diff --git a/surfsense_local/frontend/translations/pt-BR.json b/surfsense_local/frontend/translations/pt-BR.json
index dcf3f85433..6c45d1e466 100644
--- a/surfsense_local/frontend/translations/pt-BR.json
+++ b/surfsense_local/frontend/translations/pt-BR.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "Escolha o idioma que o SurfSense usa neste dispositivo.",
   "settings_general_language_system_label": "Igual ao sistema",
   "settings_general_language_title": "Idioma",
+  "settings_general_reduce_motion_body": "Mantenha as animações no mínimo neste dispositivo.",
+  "settings_general_reduce_motion_off_label": "Desativado",
+  "settings_general_reduce_motion_on_label": "Ativado",
+  "settings_general_reduce_motion_system_label": "Igual ao sistema",
+  "settings_general_reduce_motion_title": "Reduzir movimento",
   "settings_general_title": "Geral",
   "settings_image_edit_models_body": "O modelo que edita imagens. Execute um neste computador ou use um de um servidor.",
   "settings_image_edit_models_title": "Modelos de edição de imagem",
diff --git a/surfsense_local/frontend/translations/ru.json b/surfsense_local/frontend/translations/ru.json
index e15d75cec4..73c4e55ba7 100644
--- a/surfsense_local/frontend/translations/ru.json
+++ b/surfsense_local/frontend/translations/ru.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "Выберите язык, который SurfSense использует на этом устройстве.",
   "settings_general_language_system_label": "Как в системе",
   "settings_general_language_title": "Язык",
+  "settings_general_reduce_motion_body": "Сведите анимацию к минимуму на этом устройстве.",
+  "settings_general_reduce_motion_off_label": "Выключено",
+  "settings_general_reduce_motion_on_label": "Включено",
+  "settings_general_reduce_motion_system_label": "Как в системе",
+  "settings_general_reduce_motion_title": "Уменьшить движение",
   "settings_general_title": "Общие",
   "settings_image_edit_models_body": "Модель, которая редактирует изображения. Запустите её на этом компьютере или используйте модель с сервера.",
   "settings_image_edit_models_title": "Модели редактирования изображений",
diff --git a/surfsense_local/frontend/translations/zh-CN.json b/surfsense_local/frontend/translations/zh-CN.json
index dc586b4fc6..b7b5405605 100644
--- a/surfsense_local/frontend/translations/zh-CN.json
+++ b/surfsense_local/frontend/translations/zh-CN.json
@@ -1257,6 +1257,11 @@
   "settings_general_language_body": "选择 SurfSense 在这台设备上使用的语言。",
   "settings_general_language_system_label": "跟随系统",
   "settings_general_language_title": "语言",
+  "settings_general_reduce_motion_body": "在这台设备上尽量减少动画。",
+  "settings_general_reduce_motion_off_label": "关闭",
+  "settings_general_reduce_motion_on_label": "开启",
+  "settings_general_reduce_motion_system_label": "跟随系统",
+  "settings_general_reduce_motion_title": "减弱动态效果",
   "settings_general_title": "通用",
   "settings_image_edit_models_body": "编辑图像的模型。可以在这台电脑上运行一个，也可以使用服务器上的。",
   "settings_image_edit_models_title": "图像编辑模型",
diff --git a/www/content/docs/v2/docker/differences.mdx b/www/content/docs/v2/docker/differences.mdx
index 8e9e3e54e8..7f729f483f 100644
--- a/www/content/docs/v2/docker/differences.mdx
+++ b/www/content/docs/v2/docker/differences.mdx
@@ -57,7 +57,7 @@ In the second, type `/login` and sign in with your Claude account. Claude Code a
 ### The interface
 
 - The right panel's toggle sits on the Studio heading's line, or on the chat header while the panel is hidden. The desktop has it in the title bar.
-- **Match system** follows the browser's languages, and the theme follows the browser.
+- **Match system** follows the browser's languages, and the theme and **Reduce motion** follow the browser. Each browser keeps its own **Reduce motion** choice.
 - About shows the image's version. The session log and the crash notice are desktop-only.
 
 ### License

From df7ad1924b06f8807741ed55c397dea87486d6a1 Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 16:15:29 +0530
Subject: [PATCH 10/12] chore(local): remove stale motion and thread mode
 leftovers

Drops the unused motion-safe variant, points a comment at the startup
reveal's real file, and corrects the docs and a docstring that no
longer matched the code.
---
 docs/architecture/chat.md                              | 2 +-
 docs/architecture/overview.md                          | 2 +-
 surfsense_local/backend/modules/agent/engine_choice.py | 2 +-
 surfsense_local/frontend/src/index.css                 | 3 +--
 4 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/docs/architecture/chat.md b/docs/architecture/chat.md
index 1731b62e79..918538f8c6 100644
--- a/docs/architecture/chat.md
+++ b/docs/architecture/chat.md
@@ -262,7 +262,7 @@ Interface strings are `dictation_<surface>_<purpose>`, and a server code is `dic
 - The composer follows assistant-ui's keyboard behaviour.
 - While the turn waits, the thinking header's `role="status"` region says "Thinking" or the reading quarter. A second region (`reply-announcer.tsx`) stays silent until it says "Reply finished" once, for a reply it watched run to completion. It is keyed by message, because messages render by index and an instance would otherwise watch one thread's reply and announce another's. Tokens are never fed to it, since a polite region read on every delta would interrupt without pause; a reply loaded from history, a turn that ended with no answer, and one that failed or was stopped after text arrived say nothing; its alert, or the stop, already does. A stopped reply is marked `incomplete` with reason `cancelled` by the runtime ([`use-chat-runtime.ts`](../../surfsense_local/frontend/src/features/chat/use-chat-runtime.ts)), since assistant-ui would otherwise call it complete once the run ends. A queued reply's header region says "Waiting for another reply" with its place. The Chats button's count is in its accessible name ("Chats, 2 replies being written"), and an unread row carries "New reply" as screen-reader text.
 - The conversation has a heading, the thread's title, hidden visually because the visible title is the rename button. Focus does not move to it: on every thread switch the composer keeps the focus assistant-ui gives it, so someone can switch and type, and the composer is described by the heading (`aria-describedby`), so a screen reader names the conversation it writes into. A new chat has no heading until it is a thread.
-- The startup fade and the typed-in thread title respect `prefers-reduced-motion`.
+- The startup fade and the typed-in thread title follow Settings › General's Reduce motion, which matches the system by default ([overview](overview.md#frontend)).
 - The mic carries `aria-pressed` and a name that follows its state: "Start voice input", "Stop voice input", or "Set up voice input"; its tooltip is visual only. While voice input runs, a `role="status"` line in the strip says "Starting the microphone…", "Listening", "Loading the speech recognition model…", "Transcribing…" or the 30-second warning; a problem is a `role="alert"`, the two limits a `role="status"`. Under `prefers-reduced-motion` the level bars give way to the still dot beside them.
 
 ## Agent threads
diff --git a/docs/architecture/overview.md b/docs/architecture/overview.md
index 152cd08d7e..53d45d7fd3 100644
--- a/docs/architecture/overview.md
+++ b/docs/architecture/overview.md
@@ -111,7 +111,7 @@ The rules that keep the processes out of each other's way:
 - TanStack Query holds most server state: threads and messages, the citation panel, the model catalog and settings, the Studio formats and artifact list, and an open artifact with what its viewer reads. An `artifacts` event invalidates the list and every open artifact under one key prefix ([`features/studio/query-keys.ts`](../../surfsense_local/frontend/src/features/studio/query-keys.ts)), so a regenerated artifact that is open shows its new generation. The sources list is component state, reloaded on the events above.
 - There is no router. [`app/app-bootstrap.tsx`](../../surfsense_local/frontend/src/app/app-bootstrap.tsx) asks `GET /llm/onboarding`, then renders either onboarding ([`local-models/selection.md`](local-models/selection.md)) or the lazily loaded dashboard: the workspace rail, threads and sources on the left, the conversation in the middle, and Studio, artifacts and the citation panel on the right. The left and right columns resize by dragging their inner edge, and the conversation keeps at least 520 px ([studio](studio.md#the-studio-panel)).
 - While startup waits on the API, [`features/startup/`](../../surfsense_local/frontend/src/features/startup/) draws the dashboard's constant parts, with nothing to press: the columns at their saved widths, the "SurfSense" and "Studio" headings, and either a new chat's box or an open thread's message skeleton, as the last view was; a thread's composer arrives with its messages. The new chat's box takes typing but sends nothing, and the text goes to the dashboard's composer through a store `AppBootstrap` owns and provides to both (`startup-input`). The composer sets the text with `composer.setText`, then, once it shows, places the caret and tells the `@` trigger where it is through assistant-ui's trigger context (`resource.setCursorPosition`): the list tracks its own cursor, apart from the box's. Startup waits for an input method's composition to end first. One `@` rule (`mention-matcher.ts`), given to assistant-ui as its matcher, decides in both places whether the caret is in a mention. When new chats open in Agentic, which the dashboard stores for the next launch, `@` in the box shows a skeleton of the plugin list, which startup starts loading early. The dashboard draws the same parts with the same components (`composer-surface.tsx`, `conversation-scroll.tsx`, `mention-list.tsx`, rendered around assistant-ui's parts with Base UI's `useRender`), and both text boxes grow with the same autosizing textarea (`react-textarea-autosize`, which assistant-ui's input is built on), so nothing moves when it takes over; its contents fade in over 200 ms, or appear at once if the wait was under 150 ms or motion is reduced. The frame is drawn only when onboarding was last seen done on the same database, which Electron names by the database file's identity (`window.surfsense.databaseIdentity()`), so a database made since, a first launch and onboarding get the plain background. Docker has no bridge and always gets the plain background. A test walks the frame's imports with TypeScript's parser and keeps them off the chat runtime and every feature's `api.ts`, so it stays in the first chunk. No test measures the two frames' layout in a browser yet.
-- Settings › General's Reduce motion is Match system, On or Off, kept in the page's `localStorage`, so the desktop and Docker behave alike. [`features/motion-preference/`](../../surfsense_local/frontend/src/features/motion-preference/) turns it and the OS setting into one `reduce-motion` class on the document root, and `index.html` sets the class before the first frame. `motion-reduce:` and `motion-safe:` read that class, script reads `motionIsReduced()`, and one rule in `index.css` stops every other transition and animation except spinners, skeletons and scroll fades. Off can't reach the OS-only rules that sonner and shadcn's shimmer ship, so with the OS asking to reduce motion, toasts and the shimmer stay still.
+- Settings › General's Reduce motion is Match system, On or Off, kept in the page's `localStorage`, so the desktop and Docker behave alike. [`features/motion-preference/`](../../surfsense_local/frontend/src/features/motion-preference/) turns it and the OS setting into one `reduce-motion` class on the document root, and `index.html` sets the class before the first frame. `motion-reduce:` reads that class, script reads `motionIsReduced()`, and one rule in `index.css` stops every other transition and animation except spinners, skeletons and scroll fades. Off can't reach the OS-only rules that sonner and shadcn's shimmer ship, so with the OS asking to reduce motion, toasts and the shimmer stay still.
 - Code is grouped by feature under `src/features/`. Each HTTP contract lives in one feature's `api.ts`, next to its hooks and components, and other features import it from there.
 
 ## Codemap
diff --git a/surfsense_local/backend/modules/agent/engine_choice.py b/surfsense_local/backend/modules/agent/engine_choice.py
index 77affe5cde..11ed390c40 100644
--- a/surfsense_local/backend/modules/agent/engine_choice.py
+++ b/surfsense_local/backend/modules/agent/engine_choice.py
@@ -74,7 +74,7 @@ class NewThreadMode:
 async def new_thread_mode(
     session: Session, requested: ChatMode | None
 ) -> NewThreadMode:
-    """The mode a new thread opens in, and the model it was decided on.
+    """The mode a new thread opens in.
 
     Asked for Agentic and refused, it raises; with nothing asked, as older
     clients send, the default falls back to Basic where Agentic cannot run.
diff --git a/surfsense_local/frontend/src/index.css b/surfsense_local/frontend/src/index.css
index 0bea125bdf..0bd37cb28b 100644
--- a/surfsense_local/frontend/src/index.css
+++ b/surfsense_local/frontend/src/index.css
@@ -13,7 +13,6 @@
 @custom-variant dark (&:is(.dark *));
 /* Settings › General decides, not only the OS (`features/motion-preference`). */
 @custom-variant motion-reduce (&:is(.reduce-motion *));
-@custom-variant motion-safe (&:not(.reduce-motion *));
 
 @theme inline {
   --font-heading: var(--font-sans);
@@ -422,7 +421,7 @@
 }
 
 /* Startup: the dashboard's contents fade in where the loading frame already
- * drew their containers (`features/startup/startup-reveal.tsx`). The
+ * drew their containers (`features/startup/use-startup-reveal.ts`). The
  * children fade, never the container, so a drawn box such as the composer's
  * stays put. Opacity only: nothing moves out of the frame's place. */
 @keyframes ss-reveal-in {

From 4485896bf3ebffca9e154287e5d52eb46f18f38a Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 16:27:58 +0530
Subject: [PATCH 11/12] feat(local): ask only for the chat and embedding models
 in onboarding

Setup now runs welcome, privacy, chat, embedding. Image, image editing,
audio and video models are chosen in Settings, and the feature that needs
one says so. The per-slot step tables go with those steps.
---
 docs/architecture/chatgpt-subscription.md     |   2 +-
 docs/architecture/connections.md              |   2 +-
 docs/architecture/local-models/catalog.md     |   2 +-
 docs/architecture/local-models/selection.md   |  32 +-
 docs/proposals/embedding-model-choice.md      |   2 +-
 .../onboarding/model-step/kinds/slot.ts       |  16 -
 .../model-step/kinds/step-deletes.ts          |  31 --
 .../onboarding/model-step/kinds/step-kind.ts  |  11 +-
 .../model-step/kinds/step-models-table.ts     |  14 +-
 .../model-step/kinds/step-models.ts           |  19 +-
 .../model-step/kinds/use-chat-step.ts         |  32 ++
 .../model-step/kinds/use-slot-models.ts       | 103 ------
 .../model-step/kinds/use-slot-step.ts         |  28 --
 .../onboarding/model-step/model-step.tsx      |  61 ++--
 .../model-step/server/server-option.tsx       |   2 +-
 .../model-step/server/server-path.tsx         |  10 +-
 .../onboarding/model-step/step-copy.ts        |  88 +----
 .../onboarding/onboarding-page.test.tsx       | 318 +-----------------
 .../features/onboarding/onboarding-page.tsx   |  11 +-
 .../onboarding/use-finish-onboarding.ts       |   5 +-
 surfsense_local/frontend/translations/de.json |  13 -
 surfsense_local/frontend/translations/en.json |  13 -
 surfsense_local/frontend/translations/es.json |  13 -
 surfsense_local/frontend/translations/fr.json |  13 -
 surfsense_local/frontend/translations/hi.json |  13 -
 surfsense_local/frontend/translations/ja.json |  13 -
 surfsense_local/frontend/translations/ko.json |  13 -
 .../frontend/translations/pt-BR.json          |  13 -
 surfsense_local/frontend/translations/ru.json |  13 -
 .../frontend/translations/zh-CN.json          |  13 -
 www/content/docs/v2/models/api-providers.mdx  |   2 +-
 www/content/docs/v2/models/index.mdx          |   2 +-
 .../docs/v2/models/other-local-servers.mdx    |   2 +-
 www/content/docs/v2/models/other-models.mdx   |   2 +-
 34 files changed, 130 insertions(+), 797 deletions(-)
 delete mode 100644 surfsense_local/frontend/src/features/onboarding/model-step/kinds/slot.ts
 delete mode 100644 surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-deletes.ts
 create mode 100644 surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-chat-step.ts
 delete mode 100644 surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-models.ts
 delete mode 100644 surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-step.ts

diff --git a/docs/architecture/chatgpt-subscription.md b/docs/architecture/chatgpt-subscription.md
index 803c2b9e98..bb66ce7c70 100644
--- a/docs/architecture/chatgpt-subscription.md
+++ b/docs/architecture/chatgpt-subscription.md
@@ -83,7 +83,7 @@ The OpenAI catalog records the plan's models as calling tools, so Agentic mode i
 
 ## Frontend
 
-Every model list reads `serves` rather than deciding: the Settings sections, their empty states and onboarding's server steps leave a ChatGPT connection out of every section but chat. A new ChatGPT connection is made only through **Use your ChatGPT plan**, below. The connection form lists **ChatGPT subscription** only while editing a ChatGPT connection, so a server's provider list never offers it. Editing one hides the URL and key fields, shows the account, and offers **Sign in again**, with **Open the sign-in page again** and Cancel while it waits, **Sign out**, and **Save changes** for a new name. Its provider cannot be changed. The model groups show the account's email in place of the URL.
+Every model list reads `serves` rather than deciding: the Settings sections and their empty states leave a ChatGPT connection out of every section but chat. A new ChatGPT connection is made only through **Use your ChatGPT plan**, below. The connection form lists **ChatGPT subscription** only while editing a ChatGPT connection, so a server's provider list never offers it. Editing one hides the URL and key fields, shows the account, and offers **Sign in again**, with **Open the sign-in page again** and Cancel while it waits, **Sign out**, and **Save changes** for a new name. Its provider cannot be changed. The model groups show the account's email in place of the URL.
 
 **Use your ChatGPT plan** sits below **Use a server** on Settings' **Add model** page and on onboarding's text step, wherever the sign-in's `serves` includes the slot ([`chatgpt-option.tsx`](../../surfsense_local/frontend/src/features/models/remote/connections/chatgpt/chatgpt-option.tsx)). One click signs in as **ChatGPT subscription**, with the dialog's progress in the row and Cancel in place of the button, then opens the new connection's models as connecting a server does. A plan already connected reads "Connected as" its email and offers **Show models** instead.
 
diff --git a/docs/architecture/connections.md b/docs/architecture/connections.md
index 5d74043454..b8392965b3 100644
--- a/docs/architecture/connections.md
+++ b/docs/architecture/connections.md
@@ -180,7 +180,7 @@ Keys are protected by envelope encryption ([ADR 0018](../adr/0018-keychain-envel
 
 - Each model section in Settings shows, as a group under the local models, every connection whose `serves` includes that section's slot. `serves` comes from one backend rule ([`serves.py`](../../surfsense_local/backend/modules/llm/connections/serves.py)): every slot but `speech_to_text` for a key connection, `text_gen` alone for a ChatGPT one. The section's empty state counts only those connections. A group loads its models only when opened, so a slow or failed endpoint does not hold up the others, and lists only those whose `selectable_for` includes that section's slot; the model in use, if it comes from that server, stays shown above the list whether the group is open or closed. Open, its list scrolls inside its own capped-height area rather than lengthening the page. A model's row shows its type chips, and a Vision chip when `reads_images` is true. A model is assigned after an optional test; an exact ID the listing lacks can be typed in. A new connection is added from **Use a server** on the section's **Add model** page. Adding and editing a connection happen in a dialog over the page. Saving a new connection returns to the list with its group open on its models, since choosing one is why it was added; saving an edit only refreshes.
 - Edit and Disconnect sit on each group. Edit opens the same form in that dialog. A connection can fill every slot but `speech_to_text`, so Disconnect names each model it will clear, Chat, Image or both, whichever section it is disconnected from.
-- Onboarding's model steps use the same groups and the same dialog; with nothing connected yet, their **Connect** opens the dialog directly, and saving opens the server page on the new connection's models.
+- Onboarding's chat step uses the same groups and the same dialog; with nothing connected yet, its **Connect** opens the dialog directly, and saving opens the server page on the new connection's models.
 - The connection form picks a provider from the remote manifest, through `GET /llm/catalog/remote`, or "Local or custom server". A ready provider fills its URL, which stays editable so a proxy in front of it still works; a provider that needs account details asks for each field and builds the URL from its template; a provider that needs a URL leaves it to the user; an unreachable one is listed, disabled, with its reason. A provider that takes no key, a loopback server, hides the key field. "Local or custom server" leaves the URL to the user, since its port is whatever its owner set, with `http://localhost:11434/v1` as placeholder text only; a loopback server the manifest lists, such as LM Studio, fills the manifest's URL like any ready provider. The save sends `catalog_provider`.
 
 ## Known gaps
diff --git a/docs/architecture/local-models/catalog.md b/docs/architecture/local-models/catalog.md
index 9ecdf8932e..52c5974441 100644
--- a/docs/architecture/local-models/catalog.md
+++ b/docs/architecture/local-models/catalog.md
@@ -506,7 +506,7 @@ counted rather than decoded
 
 `POST /llm/installs` takes `{"catalog_id": "...", "select": true}`, and
 optionally the `model_type` that `select` fills, the engine's first type when
-absent, so onboarding's editing step installs straight into `image_edit`. Nothing
+absent, so a speech model's download installs straight into `speech_to_text`. Nothing
 else: no repo, file, URL, path or quantization, so the renderer cannot name an
 arbitrary download. A curated build's id is minted once per process and keyed on
 its repo and weights path; a searched build's is a ticket holding the whole build
diff --git a/docs/architecture/local-models/selection.md b/docs/architecture/local-models/selection.md
index 36e49a8174..f2c1fe5347 100644
--- a/docs/architecture/local-models/selection.md
+++ b/docs/architecture/local-models/selection.md
@@ -202,13 +202,15 @@ route. Only the onboarding page's last step does, once a chat model is
 persisted. Once the marker exists the app never shows onboarding again, and a
 missing selection is fixed from Settings' Chat section.
 
-The onboarding page opens on a welcome screen, then six steps: chat, image, image editing, audio, video and search model. The welcome is not counted as a step, but it is part of onboarding and gated by the same marker, so it is never shown again once onboarding is done.
+The onboarding page opens on a welcome screen, then the privacy choices ([usage statistics](../usage-statistics.md#consent)), then two model steps: chat and search model. The welcome is not counted as a step, but it is part of onboarding and gated by the same marker, so it is never shown again once onboarding is done. Image, image editing, audio and video models are not asked for: they are chosen in Settings, and the feature that needs one says so.
 
 The embedding step comes last and is not a slot, but it is the same component
-as the model steps below, with one more entry in their tables
+as the chat step below, with its own entry in their tables
 ([`model-step/`](../../../surfsense_local/frontend/src/features/onboarding/model-step/)).
-What a step does on Use is its own hook: a slot's saves the selection, and the
-embedding step's ([`use-embedding-step.ts`](../../../surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-embedding-step.ts))
+What a step does on Use is its own hook: the chat step's
+([`use-chat-step.ts`](../../../surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-chat-step.ts))
+saves the selection, and the embedding step's
+([`use-embedding-step.ts`](../../../surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-embedding-step.ts))
 only marks a choice, In use until another is used, bge-small by default and again
 if the chosen one is deleted. Its downloads install with `select: false`; a
 Hugging Face pick is labelled not tested by SurfSense; the bundled bge-small has
@@ -220,21 +222,17 @@ Finish sends the choice with the call that ends onboarding; Skip and finish, or
 Finish with the choice untouched, sends none, which means bge-small
 ([embedding](../embedding.md)).
 
-The model steps are one component for any slot
+The model steps are one component
 ([`frontend/src/features/onboarding/model-step/`](../../../surfsense_local/frontend/src/features/onboarding/model-step/)),
-built on the same hooks as Settings but with its own screens. Each lists every
-model this computer can run at once: in the chat step, models downloaded from Hugging Face first, then the curated list with its starred row first, with
+built on the same hooks as Settings but with its own screens. The chat step lists every
+model this computer can run at once: models downloaded from Hugging Face first, then the curated list with its starred row first, with
 Download, Use and Delete as in Settings, and no Delete on a model the app ships; a download's progress shows under its
-row and never moves the page. The chat step also offers Settings' Hugging Face
-search, closed until asked for; the image, image editing, audio and video steps have none, since
-sd.cpp's and audio.cpp's models are the few the catalog ships. The image editing step lists first the model chosen for images earlier when it edits too, so FLUX.2 klein is one Use away, and its downloads fill `image_edit`. A server sits one line below the list and names
-any connected earlier. Once the slot has a model, the footer names it beside
-Continue. Onboarding installs with `select: true`, so a download is also the
-choice; Settings installs with `select: false`. The chat step's Continue is
-enabled only once a chat model is selected, local or from a server. The image,
-image editing, audio and video steps are optional, and each enables its Continue or Finish only once
-its slot has a model. Each step's Skip and Continue move on to the next, image
-to image editing to audio to video. The last step's buttons read Skip and finish and Finish, and both post the marker.
+row and never moves the page. It also offers Settings' Hugging Face
+search, closed until asked for. A server sits one line below the list and names
+any connected earlier. Once a chat model is chosen, the footer names it beside
+Continue. The chat step installs with `select: true`, so a download is also the
+choice; Settings installs with `select: false`. Its Continue is
+enabled only once a chat model is selected, local or from a server, and it has no Skip. The embedding step's buttons read Skip and finish and Finish, and both post the marker.
 
 ## Resolution: local and remote
 
diff --git a/docs/proposals/embedding-model-choice.md b/docs/proposals/embedding-model-choice.md
index 82c5813e82..0194683468 100644
--- a/docs/proposals/embedding-model-choice.md
+++ b/docs/proposals/embedding-model-choice.md
@@ -257,7 +257,7 @@ bge is bundled and cannot become unreachable. Chat's check for missing files in
 
 ## Onboarding
 
-1. A new step before the chat model. It is not a model slot: [`slot.ts`](../../surfsense_local/frontend/src/features/onboarding/model-step/kinds/slot.ts) is typed to `ModelType`, and the embedder is not a selection. The step list needs a kind of step that is not one.
+1. A new step before the chat model. It is not a model slot: the embedder is not a selection, so the step list needs a kind of step that is not one ([`step-kind.ts`](../../surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-kind.ts)).
 2. Curated models come first, bge preselected, each with its download size and one line on what it is for: *"Choose Multilingual only if your documents or questions are in more than one language."* Hugging Face search and remote models sit below.
 3. A local pick not on disk resolves first (size, whether it is on disk, any error), then downloads with the hash check. A remote pick connects. Then the probe and the sanity check run, and the pick gets its label. The click is a user action, so the egress dialog may appear once.
 4. Before the user continues, the step says what cannot be undone: that the model is fixed for this library, what its label means, and for remote, where every document goes. A Hugging Face pick shows its download size with one line: *"Larger models make adding documents slower and use more memory, and this choice can't be changed later."*
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/slot.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/slot.ts
deleted file mode 100644
index 65aea3da68..0000000000
--- a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/slot.ts
+++ /dev/null
@@ -1,16 +0,0 @@
-import type { ModelType } from "@/features/models/model-type"
-
-/** The slots onboarding asks for. The others are chosen later, in Settings. */
-export type OnboardingSlot = Extract<
-  ModelType,
-  "text_gen" | "image_gen" | "image_edit" | "video_gen" | "audio_gen"
->
-
-/** The local runtime that runs each slot's downloads. */
-export const LOCAL_PROVIDER: Record<OnboardingSlot, string> = {
-  text_gen: "llamacpp",
-  image_gen: "sdcpp",
-  image_edit: "sdcpp",
-  video_gen: "sdcpp",
-  audio_gen: "audiocpp",
-}
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-deletes.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-deletes.ts
deleted file mode 100644
index 13d1bb1eb0..0000000000
--- a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-deletes.ts
+++ /dev/null
@@ -1,31 +0,0 @@
-import { useDeleteLocalAudioModel } from "@/features/models/local/audio/use-delete-local-audio-model"
-import { useDeleteLocalChatModel } from "@/features/models/local/chat/use-delete-local-chat-model"
-import { useDeleteLocalImageModel } from "@/features/models/local/image/use-delete-local-image-model"
-
-import type { OnboardingStepKind } from "./step-kind"
-
-type SlotDelete = {
-  mutateAsync: (installedAs: string) => Promise<unknown>
-  isPending: boolean
-}
-
-// No callback on clearing the chat slot: the step reads the selection itself,
-// so Continue disables on its own when the model in use is deleted.
-function useDeleteChat(): SlotDelete {
-  return useDeleteLocalChatModel()
-}
-
-// An embedder is deleted like any local model; the API refuses the shipped one.
-function useDeleteEmbedder(): SlotDelete {
-  return useDeleteLocalChatModel()
-}
-
-/** Settings' delete hooks, one per step, picked once per step. */
-export const stepDeletes: Record<OnboardingStepKind, () => SlotDelete> = {
-  text_gen: useDeleteChat,
-  image_gen: useDeleteLocalImageModel,
-  image_edit: useDeleteLocalImageModel,
-  video_gen: useDeleteLocalImageModel,
-  audio_gen: useDeleteLocalAudioModel,
-  embedding: useDeleteEmbedder,
-}
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-kind.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-kind.ts
index e75231fb0d..ba5cbcf4d8 100644
--- a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-kind.ts
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-kind.ts
@@ -1,12 +1,5 @@
-import type { OnboardingSlot } from "./slot"
-
 /**
- * What an onboarding step chooses: a slot's model, or the library's embedding
+ * What an onboarding step chooses: the chat model, or the library's embedding
  * model, which is chosen once and is never a slot (ADR 0037).
  */
-export type OnboardingStepKind = OnboardingSlot | "embedding"
-
-/** The slot a step fills, or null for the embedding step. */
-export function slotOf(kind: OnboardingStepKind): OnboardingSlot | null {
-  return kind === "embedding" ? null : kind
-}
+export type OnboardingStepKind = "text_gen" | "embedding"
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models-table.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models-table.ts
index 26ffd1b966..a7297a99b3 100644
--- a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models-table.ts
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models-table.ts
@@ -1,20 +1,10 @@
 import type { OnboardingStepKind } from "./step-kind"
 import type { StepModels } from "./step-models"
+import { useChatStep } from "./use-chat-step"
 import { useEmbeddingStep } from "./use-embedding-step"
-import { useSlotStep } from "./use-slot-step"
-
-const useTextGenStep = () => useSlotStep("text_gen")
-const useImageGenStep = () => useSlotStep("image_gen")
-const useImageEditStep = () => useSlotStep("image_edit")
-const useVideoGenStep = () => useSlotStep("video_gen")
-const useAudioGenStep = () => useSlotStep("audio_gen")
 
 /** One hook per step, picked once per step: the steps differ only here. */
 export const stepModels: Record<OnboardingStepKind, () => StepModels> = {
-  text_gen: useTextGenStep,
-  image_gen: useImageGenStep,
-  image_edit: useImageEditStep,
-  video_gen: useVideoGenStep,
-  audio_gen: useAudioGenStep,
+  text_gen: useChatStep,
   embedding: useEmbeddingStep,
 }
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models.ts
index f6a5d757c9..51d66fbdb2 100644
--- a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models.ts
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/step-models.ts
@@ -1,18 +1,25 @@
+import type { LocalRow } from "@/features/models/local/chat/api"
 import type { ModelType } from "@/features/models/model-type"
-
-import type { SlotModels } from "./use-slot-models"
+import type { InUse } from "@/features/models/your-models/your-model-row"
 
 /**
- * A step's models and how it chooses one: the seam between a slot, whose Use
- * saves a selection, and the embedding step, whose Use only marks a choice
+ * A step's models and how it chooses one: the seam between the chat step, whose
+ * Use saves a selection, and the embedding step, whose Use only marks a choice
  * that Finish sends.
  */
-export type StepModels = SlotModels & {
+export type StepModels = {
+  /** The step's local rows: tested ones and anything already on disk. */
+  rows: LocalRow[]
+  /** What the rows were priced against, where the engine prices them at all. */
+  hardware: string[] | null
+  inUse: InUse | null
+  isPending: boolean
+  error: Error | null
   /** Choose an installed model by the name it runs under. */
   use: (name: string) => void
   choosing: boolean
   /** How a download from this step installs. */
   install: { select: boolean; modelType?: ModelType }
-  /** What finishing onboarding from this step sends; null for a slot. */
+  /** What finishing onboarding from this step sends; null for the chat step. */
   value: string | null
 }
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-chat-step.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-chat-step.ts
new file mode 100644
index 0000000000..2c81dffa09
--- /dev/null
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-chat-step.ts
@@ -0,0 +1,32 @@
+import { describeHardware } from "@/features/models/local/chat/describe-hardware"
+import { useLocalChatCatalog } from "@/features/models/local/chat/use-local-chat-catalog"
+import { toastUseFailure } from "@/features/models/selection/toast-use-failure"
+import { useSelect } from "@/features/models/selection/use-selection"
+import { useChatModels } from "@/features/models/your-models/use-chat-models"
+
+import type { StepModels } from "./step-models"
+
+/** The chat step: Use saves the selection, and a download becomes it. */
+export function useChatStep(): StepModels {
+  const models = useChatModels()
+  const catalog = useLocalChatCatalog()
+  const select = useSelect("text_gen")
+  return {
+    rows: (catalog.data?.rows ?? []).filter((row) => row.engine === "llamacpp"),
+    hardware: catalog.data?.budget
+      ? describeHardware(catalog.data.budget, catalog.data.gpu_status)
+      : null,
+    inUse: models.inUse,
+    isPending: models.isPending,
+    error: models.error,
+    use: (name) =>
+      void select
+        .mutateAsync({
+          target: { provider: "llamacpp", connection_id: null, name },
+        })
+        .catch(toastUseFailure),
+    choosing: select.isPending,
+    install: { select: true, modelType: "text_gen" },
+    value: null,
+  }
+}
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-models.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-models.ts
deleted file mode 100644
index ad6d767f97..0000000000
--- a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-models.ts
+++ /dev/null
@@ -1,103 +0,0 @@
-import type { LocalRow } from "@/features/models/local/chat/api"
-import { describeHardware } from "@/features/models/local/chat/describe-hardware"
-import { useLocalChatCatalog } from "@/features/models/local/chat/use-local-chat-catalog"
-import { useLocalImageCatalog } from "@/features/models/local/image/use-local-image-catalog"
-import { useAudioModels } from "@/features/models/your-models/use-audio-models"
-import { useChatModels } from "@/features/models/your-models/use-chat-models"
-import { useImageModels } from "@/features/models/your-models/use-image-models"
-import type { InUse } from "@/features/models/your-models/your-model-row"
-
-import type { OnboardingSlot } from "./slot"
-
-export type SlotModels = {
-  /** The slot's local rows: tested ones and anything already on disk. */
-  rows: LocalRow[]
-  /** What the rows were priced against, where the engine prices them at all. */
-  hardware: string[] | null
-  inUse: InUse | null
-  isPending: boolean
-  error: Error | null
-}
-
-function useChatSlot(): SlotModels {
-  const models = useChatModels()
-  const catalog = useLocalChatCatalog()
-  return {
-    rows: (catalog.data?.rows ?? []).filter((row) => row.engine === "llamacpp"),
-    hardware: catalog.data?.budget
-      ? describeHardware(catalog.data.budget, catalog.data.gpu_status)
-      : null,
-    inUse: models.inUse,
-    isPending: models.isPending,
-    error: models.error,
-  }
-}
-
-function useImageSlot(): SlotModels {
-  const models = useImageModels()
-  const catalog = useLocalImageCatalog()
-  return {
-    rows: catalog.data ?? [],
-    // sd.cpp has no fit estimate, so there is nothing to say about the machine.
-    hardware: null,
-    inUse: models.inUse,
-    isPending: models.isPending,
-    error: models.error,
-  }
-}
-
-function useImageEditSlot(): SlotModels {
-  const models = useImageModels("image_edit")
-  const catalog = useLocalImageCatalog("image_edit")
-  const rows = catalog.data ?? []
-  // The image model chosen earlier leads when it edits too: one Use,
-  // nothing more to download.
-  const forImages = (row: LocalRow) =>
-    row.builds.some((build) => build.selected_for?.includes("image_gen"))
-  return {
-    rows: [...rows.filter(forImages), ...rows.filter((row) => !forImages(row))],
-    hardware: null,
-    inUse: models.inUse,
-    isPending: models.isPending,
-    error: models.error,
-  }
-}
-
-function useVideoSlot(): SlotModels {
-  const models = useImageModels("video_gen")
-  const catalog = useLocalImageCatalog("video_gen")
-  return {
-    rows: catalog.data ?? [],
-    // sd.cpp has no fit estimate for video either.
-    hardware: null,
-    inUse: models.inUse,
-    isPending: models.isPending,
-    error: models.error,
-  }
-}
-
-function useAudioSlot(): SlotModels {
-  const models = useAudioModels()
-  // The whole catalog, not audio's own view of it: the list reads chat's rows.
-  const catalog = useLocalChatCatalog()
-  return {
-    // Without voicing figures Settings lists no row either, so neither does this.
-    rows: (catalog.data?.rows ?? []).filter(
-      (row) => row.engine === "audiocpp" && row.voicing
-    ),
-    // audio.cpp has no fit estimate either.
-    hardware: null,
-    inUse: models.inUse,
-    isPending: models.isPending,
-    error: models.error,
-  }
-}
-
-/** One hook per slot, picked once per step, so the settings hooks stay the source. */
-export const slotModels: Record<OnboardingSlot, () => SlotModels> = {
-  text_gen: useChatSlot,
-  image_gen: useImageSlot,
-  image_edit: useImageEditSlot,
-  video_gen: useVideoSlot,
-  audio_gen: useAudioSlot,
-}
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-step.ts b/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-step.ts
deleted file mode 100644
index 5452608090..0000000000
--- a/surfsense_local/frontend/src/features/onboarding/model-step/kinds/use-slot-step.ts
+++ /dev/null
@@ -1,28 +0,0 @@
-import { toastUseFailure } from "@/features/models/selection/toast-use-failure"
-import { useSelect } from "@/features/models/selection/use-selection"
-
-import { LOCAL_PROVIDER, type OnboardingSlot } from "./slot"
-import type { StepModels } from "./step-models"
-import { slotModels } from "./use-slot-models"
-
-/** A slot's step: Use saves the selection, and a download becomes it. */
-export function useSlotStep(slot: OnboardingSlot): StepModels {
-  const models = slotModels[slot]()
-  const select = useSelect(slot)
-  return {
-    ...models,
-    use: (name) =>
-      void select
-        .mutateAsync({
-          target: {
-            provider: LOCAL_PROVIDER[slot],
-            connection_id: null,
-            name,
-          },
-        })
-        .catch(toastUseFailure),
-    choosing: select.isPending,
-    install: { select: true, modelType: slot },
-    value: null,
-  }
-}
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/model-step.tsx b/surfsense_local/frontend/src/features/onboarding/model-step/model-step.tsx
index 5eae30e0c4..0173d72e16 100644
--- a/surfsense_local/frontend/src/features/onboarding/model-step/model-step.tsx
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/model-step.tsx
@@ -22,6 +22,7 @@ import { ScrollFade } from "@/components/ui/scroll-fade"
 import { Separator } from "@/components/ui/separator"
 import { Spinner } from "@/components/ui/spinner"
 import type { LocalBuild, LocalRow } from "@/features/models/local/chat/api"
+import { useDeleteLocalChatModel } from "@/features/models/local/chat/use-delete-local-chat-model"
 import { useInstall } from "@/features/models/local/installs/use-install"
 import { ChatGPTOption } from "@/features/models/remote/connections/chatgpt/chatgpt-option"
 import { ClaudeOption } from "@/features/models/remote/connections/claude/claude-option"
@@ -39,14 +40,13 @@ import { ModelReady } from "./model-ready"
 import { STEP_COPY } from "./step-copy"
 import { ServerOption } from "./server/server-option"
 import { ServerPath } from "./server/server-path"
-import { slotOf, type OnboardingStepKind } from "./kinds/step-kind"
+import type { OnboardingStepKind } from "./kinds/step-kind"
 import { stepModels } from "./kinds/step-models-table"
-import { stepDeletes } from "./kinds/step-deletes"
 
 /**
- * One onboarding step, for a slot or the embedding model; what differs is in
+ * One onboarding step, for the chat or the embedding model; what differs is in
  * `kinds/` and `step-copy.ts`. Every local model is listed at once, the
- * recommended one first, with a server one line below for a slot: nothing is
+ * recommended one first, with a server one line below for chat: nothing is
  * hidden behind a click. The step is done once it has a model.
  */
 export function ModelStep({
@@ -65,17 +65,19 @@ export function ModelStep({
   onBack?: () => void
   /** Given what finishing from this step sends: the embedding choice, or null. */
   onNext: (value: string | null) => void
-  /** Present only where the slot is optional. */
+  /** Present only where the step is optional. */
   onSkip?: () => void
 }) {
   const copy = STEP_COPY[modelType]
   const models = stepModels[modelType]()
-  // A slot's download becomes its model, so a first model takes one click.
+  // A chat download becomes the model, so a first model takes one click.
   const { jobs, installs, install, cancel } = useInstall(models.install)
-  const remove = stepDeletes[modelType]()
+  // No callback on clearing the chat slot: the step reads the selection itself,
+  // so Continue disables on its own. The API refuses to delete the shipped embedder.
+  const remove = useDeleteLocalChatModel()
   // Remote embedders are not offered yet, so that step has no server.
-  const slot = slotOf(modelType)
-  const connections = useConnectionsServing(slot)
+  const offersServer = modelType === "text_gen"
+  const connections = useConnectionsServing(offersServer ? "text_gen" : null)
   const [onServer, setOnServer] = useState(false)
   // With nothing connected yet there is no server page to show: Connect
   // opens the dialog here, and a server saved from it opens on its models.
@@ -153,11 +155,11 @@ export function ModelStep({
   const body = () => {
     if (models.error) return <OfflineState message={models.error.message} />
     if (models.isPending) return null
-    if (onServer && slot) {
+    if (onServer) {
       return (
         // Choosing a model stays here, as in Settings: its group marks it
         // "In use", and the footer says which.
-        <ServerPath modelType={slot} openServerId={openServerId} />
+        <ServerPath openServerId={openServerId} />
       )
     }
     return (
@@ -216,19 +218,17 @@ export function ModelStep({
               {copy.noLocal()}
             </p>
           )}
-          {copy.search ? (
-            <HuggingFaceSearch
-              jobs={jobs}
-              disabled={busy}
-              onInstall={downloadBuild}
-              onCancel={cancel}
-              source={copy.search.source}
-              note={copy.search.note?.()}
-            />
-          ) : null}
+          <HuggingFaceSearch
+            jobs={jobs}
+            disabled={busy}
+            onInstall={downloadBuild}
+            onCancel={cancel}
+            source={copy.search.source}
+            note={copy.search.note?.()}
+          />
         </section>
 
-        {slot ? (
+        {offersServer ? (
           <>
             <Separator />
             <ServerOption
@@ -240,7 +240,7 @@ export function ModelStep({
               }
             />
             <ChatGPTOption
-              modelType={slot}
+              modelType="text_gen"
               dense
               onConnected={(connection) => {
                 setOpenServerId(connection.id)
@@ -248,7 +248,7 @@ export function ModelStep({
               }}
             />
             <ClaudeOption
-              modelType={slot}
+              modelType="text_gen"
               dense
               onConnected={(connection) => {
                 setOpenServerId(connection.id)
@@ -331,15 +331,10 @@ export function ModelStep({
               disabled={finishing}
               onClick={onSkip}
             >
-              {last
-                ? intl.formatMessage({
-                    id: "onboarding_model_step_skip_finish_button",
-                    defaultMessage: "Skip and finish",
-                  })
-                : intl.formatMessage({
-                    id: "onboarding_model_step_skip_button",
-                    defaultMessage: "Skip",
-                  })}
+              {intl.formatMessage({
+                id: "onboarding_model_step_skip_finish_button",
+                defaultMessage: "Skip and finish",
+              })}
             </Button>
           ) : null}
           <Button
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/server/server-option.tsx b/surfsense_local/frontend/src/features/onboarding/model-step/server/server-option.tsx
index cf87b0a869..7ba6825b38 100644
--- a/surfsense_local/frontend/src/features/onboarding/model-step/server/server-option.tsx
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/server/server-option.tsx
@@ -5,7 +5,7 @@ import { intl } from "@/i18n/intl"
 
 /**
  * The other way in, kept to one line under the local list. Servers connected
- * earlier, even on the step before, are named so they are reused, not re-added.
+ * earlier are named so they are reused, not re-added.
  */
 export function ServerOption({
   connections,
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/server/server-path.tsx b/surfsense_local/frontend/src/features/onboarding/model-step/server/server-path.tsx
index 4f612f19ff..c13b497232 100644
--- a/surfsense_local/frontend/src/features/onboarding/model-step/server/server-path.tsx
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/server/server-path.tsx
@@ -8,21 +8,17 @@ import { useConnectionsServing } from "@/features/models/remote/connections/use-
 import { ServerModelPicker } from "@/features/models/remote/models/server-model-picker"
 import { intl } from "@/i18n/intl"
 
-import type { OnboardingSlot } from "../kinds/slot"
-
 /**
- * A model from a server, in Settings' own server groups. Connecting and
+ * A chat model from a server, in Settings' own server groups. Connecting and
  * editing happen in Settings' dialog; a server just added opens on its models.
  */
 export function ServerPath({
-  modelType,
   openServerId: initialOpenServerId,
 }: {
-  modelType: OnboardingSlot
   /** A server added before this page opened, whose models to show first. */
   openServerId: number | null
 }) {
-  const connections = useConnectionsServing(modelType)
+  const connections = useConnectionsServing("text_gen")
   const [dialog, setDialog] = useState<Connection | "new" | null>(null)
   const [dialogOpen, setDialogOpen] = useState(false)
   const openDialog = (target: Connection | "new") => {
@@ -38,7 +34,7 @@ export function ServerPath({
     <div className="flex flex-col gap-5">
       {servers.length ? (
         <ServerModelPicker
-          modelType={modelType}
+          modelType="text_gen"
           openServerId={openServerId}
           onEdit={openDialog}
         />
diff --git a/surfsense_local/frontend/src/features/onboarding/model-step/step-copy.ts b/surfsense_local/frontend/src/features/onboarding/model-step/step-copy.ts
index 66acc435e2..97ebcee5b3 100644
--- a/surfsense_local/frontend/src/features/onboarding/model-step/step-copy.ts
+++ b/surfsense_local/frontend/src/features/onboarding/model-step/step-copy.ts
@@ -11,8 +11,8 @@ export const STEP_COPY: Record<
     title: () => string
     description: () => string
     noLocal: () => string
-    /** Hugging Face search, GGUF unless a source is given; null for none. */
-    search: { source?: SearchSource; note?: () => string } | null
+    /** Hugging Face search, GGUF unless a source is given. */
+    search: { source?: SearchSource; note?: () => string }
     /** Said above the list, where a choice has a consequence to state. */
     notice?: () => string
   }
@@ -37,90 +37,6 @@ export const STEP_COPY: Record<
       }),
     search: {},
   },
-  image_gen: {
-    title: () =>
-      intl.formatMessage({
-        id: "onboarding_image_step_title",
-        defaultMessage: "Choose an image generation model",
-      }),
-    description: () =>
-      intl.formatMessage({
-        id: "onboarding_image_step_body",
-        defaultMessage:
-          "Creates images for you. Run one on this computer, or use one from a server.",
-      }),
-    noLocal: () =>
-      intl.formatMessage({
-        id: "onboarding_image_step_no_local_empty",
-        defaultMessage:
-          "Image models cannot run on this computer. Use a server instead.",
-      }),
-    // sd.cpp has no search: its models are the few the catalog ships.
-    search: null,
-  },
-  image_edit: {
-    title: () =>
-      intl.formatMessage({
-        id: "onboarding_image_edit_step_title",
-        defaultMessage: "Choose an image editing model",
-      }),
-    description: () =>
-      intl.formatMessage({
-        id: "onboarding_image_edit_step_body",
-        defaultMessage:
-          "Edits images for you. Run one on this computer, or use one from a server.",
-      }),
-    noLocal: () =>
-      intl.formatMessage({
-        id: "onboarding_image_edit_step_no_local_empty",
-        defaultMessage:
-          "Image editing models cannot run on this computer. Use a server instead.",
-      }),
-    // Nor has it for editing: the same few models.
-    search: null,
-  },
-  video_gen: {
-    title: () =>
-      intl.formatMessage({
-        id: "onboarding_video_step_title",
-        defaultMessage: "Choose a video generation model",
-      }),
-    description: () =>
-      intl.formatMessage({
-        id: "onboarding_video_step_body",
-        defaultMessage:
-          "Makes short video clips. Run one on this computer, or use one from a server.",
-      }),
-    noLocal: () =>
-      intl.formatMessage({
-        id: "onboarding_video_step_no_local_empty",
-        defaultMessage:
-          "Video models cannot run on this computer. Use a server instead.",
-      }),
-    // Nor for video.
-    search: null,
-  },
-  audio_gen: {
-    title: () =>
-      intl.formatMessage({
-        id: "onboarding_audio_step_title",
-        defaultMessage: "Choose an audio model",
-      }),
-    description: () =>
-      intl.formatMessage({
-        id: "onboarding_audio_step_body",
-        defaultMessage:
-          "Creates podcasts for you. Run one on this computer, or use one from a server.",
-      }),
-    noLocal: () =>
-      intl.formatMessage({
-        id: "onboarding_audio_step_no_local_empty",
-        defaultMessage:
-          "Audio models cannot run on this computer. Use a server instead.",
-      }),
-    // Nor has audio.cpp.
-    search: null,
-  },
   embedding: {
     title: () =>
       intl.formatMessage({
diff --git a/surfsense_local/frontend/src/features/onboarding/onboarding-page.test.tsx b/surfsense_local/frontend/src/features/onboarding/onboarding-page.test.tsx
index 1f764c0b65..57cd97695b 100644
--- a/surfsense_local/frontend/src/features/onboarding/onboarding-page.test.tsx
+++ b/surfsense_local/frontend/src/features/onboarding/onboarding-page.test.tsx
@@ -87,10 +87,7 @@ function row(
     default_quantization: "Q4_K_M",
     recommended: false,
     engine,
-    voicing:
-      engine === "audiocpp"
-        ? { peak_mb: 1421, voice_count: 46, languages: ["en-US"] }
-        : null,
+    voicing: null,
     lead: { quantization: "Q4_K_M", why: "recommended" },
     builds: [
       {
@@ -436,37 +433,9 @@ async function toChatStep(user: ReturnType<typeof userEvent.setup>) {
   await screen.findByRole("heading", { name: "Choose a text generation model" })
 }
 
-async function toImageStep(user: ReturnType<typeof userEvent.setup>) {
+async function toEmbeddingStep(user: ReturnType<typeof userEvent.setup>) {
   await toChatStep(user)
   await user.click(await screen.findByRole("button", { name: "Continue" }))
-  await screen.findByRole("heading", {
-    name: "Choose an image generation model",
-  })
-}
-
-async function toImageEditStep(user: ReturnType<typeof userEvent.setup>) {
-  await toImageStep(user)
-  await user.click(screen.getByRole("button", { name: "Skip" }))
-  await screen.findByRole("heading", { name: "Choose an image editing model" })
-}
-
-async function toAudioStep(user: ReturnType<typeof userEvent.setup>) {
-  await toImageEditStep(user)
-  await user.click(screen.getByRole("button", { name: "Skip" }))
-  await screen.findByRole("heading", { name: "Choose an audio model" })
-}
-
-async function toVideoStep(user: ReturnType<typeof userEvent.setup>) {
-  await toAudioStep(user)
-  await user.click(screen.getByRole("button", { name: "Skip" }))
-  await screen.findByRole("heading", {
-    name: "Choose a video generation model",
-  })
-}
-
-async function toEmbeddingStep(user: ReturnType<typeof userEvent.setup>) {
-  await toVideoStep(user)
-  await user.click(screen.getByRole("button", { name: "Skip" }))
   await screen.findByRole("heading", {
     name: "Choose an embedding model",
   })
@@ -485,7 +454,7 @@ const openRouter = {
 }
 
 describe("onboarding", () => {
-  it("walks the welcome, then six steps: chat, image, image editing, audio, video and search models", async () => {
+  it("walks the welcome, then two steps: chat and search models", async () => {
     vi.stubGlobal("fetch", backend({ selections: { ...chatChosen } }))
     const user = userEvent.setup()
     render(<OnboardingPage onComplete={() => undefined} />)
@@ -493,150 +462,29 @@ describe("onboarding", () => {
     // The welcome introduces onboarding; it is not one of its steps.
     expect(screen.queryByLabelText(/Onboarding step/)).toBeNull()
     await toChatStep(user)
-    expect(screen.getByLabelText("Onboarding step 1 of 6")).toBeTruthy()
+    expect(screen.getByLabelText("Onboarding step 1 of 2")).toBeTruthy()
     // The welcome is not somewhere to go back to.
     expect(screen.queryByRole("button", { name: "Back" })).toBeNull()
+    // Chat is required.
+    expect(screen.queryByText("Optional")).toBeNull()
+    expect(screen.queryByRole("button", { name: /^Skip/ })).toBeNull()
 
     await user.click(await screen.findByRole("button", { name: "Continue" }))
     expect(
-      await screen.findByRole("heading", {
-        name: "Choose an image generation model",
-      })
-    ).toBeTruthy()
-    expect(screen.getByLabelText("Onboarding step 2 of 6")).toBeTruthy()
-    expect(screen.getByText("Optional")).toBeTruthy()
-
-    // Skipping the image model moves on; it does not end onboarding.
-    await user.click(screen.getByRole("button", { name: "Skip" }))
-    expect(
-      await screen.findByRole("heading", {
-        name: "Choose an image editing model",
-      })
-    ).toBeTruthy()
-    expect(screen.getByLabelText("Onboarding step 3 of 6")).toBeTruthy()
-
-    await user.click(screen.getByRole("button", { name: "Skip" }))
-    expect(
-      await screen.findByRole("heading", { name: "Choose an audio model" })
-    ).toBeTruthy()
-    expect(screen.getByLabelText("Onboarding step 4 of 6")).toBeTruthy()
-
-    await user.click(screen.getByRole("button", { name: "Skip" }))
-    expect(
-      await screen.findByRole("heading", {
-        name: "Choose a video generation model",
-      })
+      await screen.findByRole("heading", { name: "Choose an embedding model" })
     ).toBeTruthy()
-    expect(screen.getByLabelText("Onboarding step 5 of 6")).toBeTruthy()
-
-    await user.click(screen.getByRole("button", { name: "Skip" }))
-    expect(
-      await screen.findByRole("heading", {
-        name: "Choose an embedding model",
-      })
-    ).toBeTruthy()
-    expect(screen.getByLabelText("Onboarding step 6 of 6")).toBeTruthy()
+    expect(screen.getByLabelText("Onboarding step 2 of 2")).toBeTruthy()
     expect(screen.getByText("Optional")).toBeTruthy()
     expect(screen.getByRole("button", { name: "Finish" })).toBeTruthy()
     // Skipping the last step ends onboarding, so it says so.
     expect(screen.getByRole("button", { name: "Skip and finish" })).toBeTruthy()
-    expect(screen.queryByRole("button", { name: "Skip" })).toBeNull()
-
-    for (const heading of [
-      "Choose a video generation model",
-      "Choose an audio model",
-      "Choose an image editing model",
-      "Choose an image generation model",
-      "Choose a text generation model",
-    ]) {
-      await user.click(screen.getByRole("button", { name: "Back" }))
-      expect(await screen.findByRole("heading", { name: heading })).toBeTruthy()
-    }
-  })
 
-  it("lists only video models on the video step, and fills the video slot", async () => {
-    const video = { selectable_for: ["video_gen"], types: ["video_gen"] }
-    const fetchMock = backend({
-      selections: { ...chatChosen },
-      rows: [
-        row("llamacpp", "Qwen3 4B", "qwen3-4b", { recommended: true }),
-        row("sdcpp", "Wan2.1 1.3B", "wan-small", video),
-        row("sdcpp", "SD 1.5", "sd15"),
-      ],
-    })
-    vi.stubGlobal("fetch", fetchMock)
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={() => undefined} />)
-    await toVideoStep(user)
-
-    expect(screen.queryByText("SD 1.5")).toBeNull()
-    await user.click(
-      await screen.findByRole("button", { name: "Download Wan2.1 1.3B Q4_K_M" })
-    )
-
-    await waitFor(() =>
-      expect(installsStarted(fetchMock)).toEqual([
-        {
-          catalog_id: "opaque-wan-small",
-          select: true,
-          model_type: "video_gen",
-        },
-      ])
-    )
-  })
-
-  it("leads the editing step with the image model when it edits too", async () => {
-    // FLUX.2 klein makes and edits images from the same files: chosen for
-    // images earlier, it is one Use away, with nothing to download.
-    const edits = {
-      selectable_for: ["image_gen", "image_edit"],
-      types: ["image_gen", "image_edit"],
-    }
-    const fetchMock = backend({
-      selections: {
-        ...chatChosen,
-        image_gen: {
-          model_type: "image_gen",
-          provider: "sdcpp",
-          connection_id: null,
-          name: "flux-klein",
-          updated_at: "2026-09-25T00:00:00Z",
-        },
-      },
-      rows: [
-        row("llamacpp", "Qwen3 4B", "qwen3-4b", { recommended: true }),
-        row("sdcpp", "LongCat Edit", "longcat-edit", edits),
-        row("sdcpp", "FLUX.2 klein", "flux-klein", edits),
-        row("sdcpp", "SD 1.5", "sd15"),
-      ],
-      onDisk: ["flux-klein"],
-    })
-    vi.stubGlobal("fetch", fetchMock)
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={() => undefined} />)
-    await toImageStep(user)
-    await user.click(await screen.findByRole("button", { name: "Continue" }))
-    await screen.findByRole("heading", {
-      name: "Choose an image editing model",
-    })
-
-    const choices = await screen.findAllByRole("listitem")
-    expect(choices[0]?.textContent).toContain("FLUX.2 klein")
-    expect(screen.queryByText("SD 1.5")).toBeNull()
-    await user.click(
-      screen.getByRole("button", { name: "Use FLUX.2 klein Q4_K_M" })
-    )
-
-    await waitFor(() => {
-      const write = fetchMock.mock.calls.find(
-        ([path, init]) =>
-          path === "/llm/selection/image_edit" && init?.method === "PUT"
-      )
-      expect(JSON.parse(String(write?.[1]?.body))).toMatchObject({
-        provider: "sdcpp",
-        name: "flux-klein",
+    await user.click(screen.getByRole("button", { name: "Back" }))
+    expect(
+      await screen.findByRole("heading", {
+        name: "Choose a text generation model",
       })
-    })
+    ).toBeTruthy()
   })
 
   it("lists every model at once, the recommended one first", async () => {
@@ -844,30 +692,6 @@ describe("onboarding", () => {
     )
   })
 
-  it("offers no search on the image, image editing, video or audio step", async () => {
-    vi.stubGlobal("fetch", backend({ selections: { ...chatChosen } }))
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={() => undefined} />)
-    const search = () =>
-      screen.queryByRole("button", { name: "Not listed? Search Hugging Face" })
-
-    await toImageStep(user)
-    expect(search()).toBeNull()
-    await user.click(screen.getByRole("button", { name: "Skip" }))
-    await screen.findByRole("heading", {
-      name: "Choose an image editing model",
-    })
-    expect(search()).toBeNull()
-    await user.click(screen.getByRole("button", { name: "Skip" }))
-    await screen.findByRole("heading", { name: "Choose an audio model" })
-    expect(search()).toBeNull()
-    await user.click(screen.getByRole("button", { name: "Skip" }))
-    await screen.findByRole("heading", {
-      name: "Choose a video generation model",
-    })
-    expect(search()).toBeNull()
-  })
-
   it("offers a ChatGPT plan below Use a server on the chat step", async () => {
     vi.stubGlobal("fetch", backend({ chatgpt: true }))
     const user = userEvent.setup()
@@ -937,7 +761,6 @@ describe("onboarding", () => {
     const user = userEvent.setup()
     render(<OnboardingPage onComplete={() => undefined} />)
     await toChatStep(user)
-    await user.click(await screen.findByRole("button", { name: "Continue" }))
 
     const server = await screen.findByRole("region", { name: "Use a server" })
     expect(server.textContent).toContain("Connected: OpenRouter")
@@ -1013,48 +836,6 @@ describe("onboarding", () => {
     )
   })
 
-  it("says when image models cannot run here", async () => {
-    vi.stubGlobal("fetch", backend({ selections: { ...chatChosen } }))
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={() => undefined} />)
-    await toImageStep(user)
-
-    expect(
-      await screen.findByText(/Image models cannot run on this computer/)
-    ).toBeTruthy()
-    expect(
-      screen.getByRole("button", { name: "Continue" }).hasAttribute("disabled")
-    ).toBe(true)
-  })
-
-  it("offers a server on the audio step, as on every other", async () => {
-    vi.stubGlobal("fetch", backend({ selections: { ...chatChosen } }))
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={() => undefined} />)
-    await toAudioStep(user)
-
-    expect(
-      await screen.findByText(
-        "Audio models cannot run on this computer. Use a server instead."
-      )
-    ).toBeTruthy()
-    expect(screen.getByRole("region", { name: "Use a server" })).toBeTruthy()
-  })
-
-  it("says when audio models cannot run here", async () => {
-    vi.stubGlobal("fetch", backend({ selections: { ...chatChosen } }))
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={() => undefined} />)
-    await toAudioStep(user)
-
-    expect(
-      await screen.findByText(/Audio models cannot run on this computer/)
-    ).toBeTruthy()
-    expect(
-      screen.getByRole("button", { name: "Continue" }).hasAttribute("disabled")
-    ).toBe(true)
-  })
-
   it("says in a toast when setup cannot finish, never above the footer", async () => {
     vi.stubGlobal(
       "fetch",
@@ -1079,7 +860,7 @@ describe("onboarding", () => {
     expect(screen.queryByText("Internal Server Error")).toBeNull()
   })
 
-  it("finishes without any optional model when every one is skipped", async () => {
+  it("finishes with the chat model when the optional step is skipped", async () => {
     const fetchMock = backend({ selections: { ...chatChosen } })
     vi.stubGlobal("fetch", fetchMock)
     const onComplete = vi.fn()
@@ -1099,70 +880,7 @@ describe("onboarding", () => {
     ).toBe(true)
   })
 
-  it("moves on once an image model is downloaded", async () => {
-    const fetchMock = backend({
-      selections: { ...chatChosen },
-      rows: [
-        row("llamacpp", "Qwen3 4B", "qwen3-4b", { recommended: true }),
-        row("sdcpp", "SDXL Turbo", "sdxl-turbo"),
-      ],
-    })
-    vi.stubGlobal("fetch", fetchMock)
-    const onComplete = vi.fn()
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={onComplete} />)
-    await toImageStep(user)
-
-    await user.click(
-      await screen.findByRole("button", { name: "Download SDXL Turbo Q4_K_M" })
-    )
-    await expectReady("SDXL Turbo")
-    const next = screen.getByRole("button", { name: "Continue" })
-    await waitFor(() => expect(next.hasAttribute("disabled")).toBe(false))
-    await user.click(next)
-
-    expect(
-      await screen.findByRole("heading", {
-        name: "Choose an image editing model",
-      })
-    ).toBeTruthy()
-    expect(onComplete).not.toHaveBeenCalled()
-  })
-
-  it("moves on to the search step once a video model is downloaded", async () => {
-    const fetchMock = backend({
-      selections: { ...chatChosen },
-      rows: [
-        row("llamacpp", "Qwen3 4B", "qwen3-4b", { recommended: true }),
-        row("sdcpp", "Wan2.1 1.3B", "wan-small", {
-          selectable_for: ["video_gen"],
-          types: ["video_gen"],
-        }),
-      ],
-    })
-    vi.stubGlobal("fetch", fetchMock)
-    const onComplete = vi.fn()
-    const user = userEvent.setup()
-    render(<OnboardingPage onComplete={onComplete} />)
-    await toVideoStep(user)
-
-    await user.click(
-      await screen.findByRole("button", { name: "Download Wan2.1 1.3B Q4_K_M" })
-    )
-    await expectReady("Wan2.1 1.3B")
-    const next = screen.getByRole("button", { name: "Continue" })
-    await waitFor(() => expect(next.hasAttribute("disabled")).toBe(false))
-    await user.click(next)
-
-    expect(
-      await screen.findByRole("heading", {
-        name: "Choose an embedding model",
-      })
-    ).toBeTruthy()
-    expect(onComplete).not.toHaveBeenCalled()
-  })
-
-  it("never finishes onboarding from a model step", async () => {
+  it("never finishes onboarding until the last step is left", async () => {
     const fetchMock = backend({ selections: { ...chatChosen } })
     vi.stubGlobal("fetch", fetchMock)
     const onComplete = vi.fn()
@@ -1448,7 +1166,7 @@ describe("the privacy choices", () => {
 
     await toPrivacyStep(user)
     // The first step like any other: counted, with no way back to the welcome.
-    expect(screen.getByLabelText("Onboarding step 1 of 7")).toBeTruthy()
+    expect(screen.getByLabelText("Onboarding step 1 of 3")).toBeTruthy()
     expect(screen.queryByRole("button", { name: "Back" })).toBeNull()
     for (const row of ["Check for updates", "Share usage statistics"])
       for (const answer of ["Yes", "No"] as const)
@@ -1672,7 +1390,7 @@ describe("the privacy choices", () => {
     await screen.findByRole("heading", {
       name: "Choose a text generation model",
     })
-    expect(screen.getByLabelText("Onboarding step 2 of 7")).toBeTruthy()
+    expect(screen.getByLabelText("Onboarding step 2 of 3")).toBeTruthy()
 
     await user.click(screen.getByRole("button", { name: "Back" }))
     await screen.findByRole("heading", { name: "Privacy choices" })
diff --git a/surfsense_local/frontend/src/features/onboarding/onboarding-page.tsx b/surfsense_local/frontend/src/features/onboarding/onboarding-page.tsx
index bd33bf19a3..9b1db11ff4 100644
--- a/surfsense_local/frontend/src/features/onboarding/onboarding-page.tsx
+++ b/surfsense_local/frontend/src/features/onboarding/onboarding-page.tsx
@@ -19,17 +19,14 @@ import { useFinishOnboarding } from "./use-finish-onboarding"
  * The model steps, in order. The privacy choices come before them as the first
  * step the progress bar counts. The welcome is still onboarding, and still gated
  * by the same marker, but it is an introduction, not a step to complete.
- * Whichever step is last finishes onboarding.
+ * Whichever step is last finishes onboarding. Every other kind of model is
+ * chosen later, in Settings.
  *
- * The embedding model comes last and is the one step that is not a slot: its
- * choice is sent with the finish that fixes it, never held across steps.
+ * The embedding model comes last and is not a slot: its choice is sent with the
+ * finish that fixes it, never held across steps.
  */
 const MODEL_STEPS = [
   "text_gen",
-  "image_gen",
-  "image_edit",
-  "audio_gen",
-  "video_gen",
   "embedding",
 ] as const satisfies readonly OnboardingStepKind[]
 
diff --git a/surfsense_local/frontend/src/features/onboarding/use-finish-onboarding.ts b/surfsense_local/frontend/src/features/onboarding/use-finish-onboarding.ts
index b6c9f54929..a59cf6a9e2 100644
--- a/surfsense_local/frontend/src/features/onboarding/use-finish-onboarding.ts
+++ b/surfsense_local/frontend/src/features/onboarding/use-finish-onboarding.ts
@@ -10,9 +10,8 @@ import { completeOnboarding } from "./api"
 
 /**
  * Marks onboarding done and hands the app its chat model. Only this may call
- * the route, and only once a chat model is chosen: the image model is optional,
- * so Skip and Finish end the same way. The search model passed in is fixed
- * here, for good; null means the one SurfSense ships.
+ * the route, and only once a chat model is chosen. The search model passed in
+ * is fixed here, for good; null means the one SurfSense ships.
  */
 export function useFinishOnboarding(
   onComplete: (selection: ModelSelection) => void
diff --git a/surfsense_local/frontend/translations/de.json b/surfsense_local/frontend/translations/de.json
index 1f05129300..5522d87316 100644
--- a/surfsense_local/frontend/translations/de.json
+++ b/surfsense_local/frontend/translations/de.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Office-Unterstützung",
   "office_support_settings_turn_on_button": "Einschalten",
   "office_support_settings_using_yours_status": "Deine eigene Installation unter {path}",
-  "onboarding_audio_step_body": "Erstellt Podcasts für dich. Führe eines auf diesem Computer aus oder nutze eines von einem Server.",
-  "onboarding_audio_step_no_local_empty": "Audiomodelle können auf diesem Computer nicht laufen. Nutze stattdessen einen Server.",
-  "onboarding_audio_step_title": "Audiomodell wählen",
   "onboarding_chat_step_body": "Antwortet dir im Chat. Führe eines auf diesem Computer aus, damit deine Chats privat bleiben, oder nutze eines von einem Server.",
   "onboarding_chat_step_no_local_empty": "Kein getestetes Modell kann auf diesem Computer laufen. Nutze stattdessen einen Server.",
   "onboarding_chat_step_title": "Modell für Texterzeugung wählen",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "Wähle zuerst ein Chatmodell",
   "onboarding_hugging_face_search_hide_button": "Hugging-Face-Suche ausblenden",
   "onboarding_hugging_face_search_show_button": "Nicht dabei? Auf Hugging Face suchen",
-  "onboarding_image_edit_step_body": "Bearbeitet Bilder für dich. Führe eines auf diesem Computer aus oder nutze eines von einem Server.",
-  "onboarding_image_edit_step_no_local_empty": "Bildbearbeitungsmodelle können auf diesem Computer nicht laufen. Nutze stattdessen einen Server.",
-  "onboarding_image_edit_step_title": "Modell für Bildbearbeitung wählen",
-  "onboarding_image_step_body": "Erstellt Bilder für dich. Führe eines auf diesem Computer aus oder nutze eines von einem Server.",
-  "onboarding_image_step_no_local_empty": "Bildmodelle können auf diesem Computer nicht laufen. Nutze stattdessen einen Server.",
-  "onboarding_image_step_title": "Modell für Bilderzeugung wählen",
   "onboarding_model_list_aria": "Modelle für diesen Computer",
   "onboarding_model_list_delete_aria": "{name} löschen",
   "onboarding_model_list_hugging_face_label": "Von Hugging Face",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "Auf diesem Computer",
   "onboarding_model_step_local_title": "Auf diesem Computer",
   "onboarding_model_step_optional_label": "Optional",
-  "onboarding_model_step_skip_button": "Überspringen",
   "onboarding_model_step_skip_finish_button": "Überspringen und fertigstellen",
   "onboarding_offline_start_body": "Starte es mit diesem Befehl:",
   "onboarding_offline_title": "Lokales Backend nicht verfügbar",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "Weiteren Server verbinden",
   "onboarding_server_path_connect_button": "Server verbinden",
   "onboarding_server_path_empty": "Noch keine Server verbunden.",
-  "onboarding_video_step_body": "Erstellt kurze Videoclips. Führe eines auf diesem Computer aus oder nutze eines von einem Server.",
-  "onboarding_video_step_no_local_empty": "Videomodelle können auf diesem Computer nicht laufen. Nutze stattdessen einen Server.",
-  "onboarding_video_step_title": "Modell für Videoerzeugung wählen",
   "onboarding_welcome_body": "Ein privates Recherche-Notizbuch, das komplett auf deinem eigenen Rechner läuft. Deine Dokumente, deine Modellschlüssel, keine Cloud, kein Konto.",
   "onboarding_welcome_start_button": "Einrichtung starten",
   "onboarding_welcome_title": "Die <accent>NotebookLM-Alternative</accent>: air-gapped und Open Source",
diff --git a/surfsense_local/frontend/translations/en.json b/surfsense_local/frontend/translations/en.json
index 03c3c8adb6..dde29648ac 100644
--- a/surfsense_local/frontend/translations/en.json
+++ b/surfsense_local/frontend/translations/en.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Office support",
   "office_support_settings_turn_on_button": "Turn on",
   "office_support_settings_using_yours_status": "Your own install at {path}",
-  "onboarding_audio_step_body": "Creates podcasts for you. Run one on this computer, or use one from a server.",
-  "onboarding_audio_step_no_local_empty": "Audio models cannot run on this computer. Use a server instead.",
-  "onboarding_audio_step_title": "Choose an audio model",
   "onboarding_chat_step_body": "Answers you in chat. Run one on this computer so your chats stay private, or use one from a server.",
   "onboarding_chat_step_no_local_empty": "No tested model can run on this computer. Use a server instead.",
   "onboarding_chat_step_title": "Choose a text generation model",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "Choose a chat model first",
   "onboarding_hugging_face_search_hide_button": "Hide Hugging Face search",
   "onboarding_hugging_face_search_show_button": "Not listed? Search Hugging Face",
-  "onboarding_image_edit_step_body": "Edits images for you. Run one on this computer, or use one from a server.",
-  "onboarding_image_edit_step_no_local_empty": "Image editing models cannot run on this computer. Use a server instead.",
-  "onboarding_image_edit_step_title": "Choose an image editing model",
-  "onboarding_image_step_body": "Creates images for you. Run one on this computer, or use one from a server.",
-  "onboarding_image_step_no_local_empty": "Image models cannot run on this computer. Use a server instead.",
-  "onboarding_image_step_title": "Choose an image generation model",
   "onboarding_model_list_aria": "Models for this computer",
   "onboarding_model_list_delete_aria": "Delete {name}",
   "onboarding_model_list_hugging_face_label": "From Hugging Face",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "On this computer",
   "onboarding_model_step_local_title": "On this computer",
   "onboarding_model_step_optional_label": "Optional",
-  "onboarding_model_step_skip_button": "Skip",
   "onboarding_model_step_skip_finish_button": "Skip and finish",
   "onboarding_offline_start_body": "Start it with this command:",
   "onboarding_offline_title": "Local backend unavailable",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "Connect another server",
   "onboarding_server_path_connect_button": "Connect a server",
   "onboarding_server_path_empty": "No servers connected yet.",
-  "onboarding_video_step_body": "Makes short video clips. Run one on this computer, or use one from a server.",
-  "onboarding_video_step_no_local_empty": "Video models cannot run on this computer. Use a server instead.",
-  "onboarding_video_step_title": "Choose a video generation model",
   "onboarding_welcome_body": "A private research notebook that runs entirely on your own machine. Your documents, your model keys, no cloud, no account.",
   "onboarding_welcome_start_button": "Start setting up",
   "onboarding_welcome_title": "Air-gapped, open source <accent>NotebookLM alternative</accent>",
diff --git a/surfsense_local/frontend/translations/es.json b/surfsense_local/frontend/translations/es.json
index 477cc1097a..50e192e7a6 100644
--- a/surfsense_local/frontend/translations/es.json
+++ b/surfsense_local/frontend/translations/es.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Compatibilidad con Office",
   "office_support_settings_turn_on_button": "Activar",
   "office_support_settings_using_yours_status": "Tu propia instalación en {path}",
-  "onboarding_audio_step_body": "Crea podcasts para ti. Ejecuta uno en este equipo o usa uno de un servidor.",
-  "onboarding_audio_step_no_local_empty": "Los modelos de audio no pueden ejecutarse en este equipo. Usa un servidor en su lugar.",
-  "onboarding_audio_step_title": "Elige un modelo de audio",
   "onboarding_chat_step_body": "Te responde en el chat. Ejecuta uno en este equipo para que tus chats sigan siendo privados, o usa uno de un servidor.",
   "onboarding_chat_step_no_local_empty": "Ningún modelo probado puede ejecutarse en este equipo. Usa un servidor en su lugar.",
   "onboarding_chat_step_title": "Elige un modelo de generación de texto",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "Elige primero un modelo de chat",
   "onboarding_hugging_face_search_hide_button": "Ocultar la búsqueda en Hugging Face",
   "onboarding_hugging_face_search_show_button": "¿No está en la lista? Busca en Hugging Face",
-  "onboarding_image_edit_step_body": "Edita imágenes para ti. Ejecuta uno en este equipo o usa uno de un servidor.",
-  "onboarding_image_edit_step_no_local_empty": "Los modelos de edición de imágenes no pueden ejecutarse en este equipo. Usa un servidor en su lugar.",
-  "onboarding_image_edit_step_title": "Elige un modelo de edición de imágenes",
-  "onboarding_image_step_body": "Crea imágenes para ti. Ejecuta uno en este equipo o usa uno de un servidor.",
-  "onboarding_image_step_no_local_empty": "Los modelos de imagen no pueden ejecutarse en este equipo. Usa un servidor en su lugar.",
-  "onboarding_image_step_title": "Elige un modelo de generación de imágenes",
   "onboarding_model_list_aria": "Modelos para este equipo",
   "onboarding_model_list_delete_aria": "Eliminar {name}",
   "onboarding_model_list_hugging_face_label": "De Hugging Face",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "En este equipo",
   "onboarding_model_step_local_title": "En este equipo",
   "onboarding_model_step_optional_label": "Opcional",
-  "onboarding_model_step_skip_button": "Omitir",
   "onboarding_model_step_skip_finish_button": "Omitir y finalizar",
   "onboarding_offline_start_body": "Inícialo con este comando:",
   "onboarding_offline_title": "El backend local no está disponible",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "Conectar otro servidor",
   "onboarding_server_path_connect_button": "Conectar un servidor",
   "onboarding_server_path_empty": "Todavía no hay servidores conectados.",
-  "onboarding_video_step_body": "Crea videoclips cortos. Ejecuta uno en este equipo o usa uno de un servidor.",
-  "onboarding_video_step_no_local_empty": "Los modelos de video no pueden ejecutarse en este equipo. Usa un servidor en su lugar.",
-  "onboarding_video_step_title": "Elige un modelo de generación de video",
   "onboarding_welcome_body": "Un cuaderno de investigación privado que funciona por completo en tu propia máquina. Tus documentos, tus claves de modelos, sin nube y sin cuenta.",
   "onboarding_welcome_start_button": "Empezar la configuración",
   "onboarding_welcome_title": "La <accent>alternativa a NotebookLM</accent>: aislada de la red y de código abierto",
diff --git a/surfsense_local/frontend/translations/fr.json b/surfsense_local/frontend/translations/fr.json
index e6580e22e0..d1a9147297 100644
--- a/surfsense_local/frontend/translations/fr.json
+++ b/surfsense_local/frontend/translations/fr.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Prise en charge d’Office",
   "office_support_settings_turn_on_button": "Activer",
   "office_support_settings_using_yours_status": "Votre propre installation dans {path}",
-  "onboarding_audio_step_body": "Crée vos podcasts. Exécutez-en un sur cet ordinateur ou utilisez-en un depuis un serveur.",
-  "onboarding_audio_step_no_local_empty": "Les modèles audio ne peuvent pas fonctionner sur cet ordinateur. Utilisez plutôt un serveur.",
-  "onboarding_audio_step_title": "Choisissez un modèle audio",
   "onboarding_chat_step_body": "Vous répond dans le chat. Exécutez-en un sur cet ordinateur pour que vos chats restent privés, ou utilisez-en un depuis un serveur.",
   "onboarding_chat_step_no_local_empty": "Aucun modèle testé ne peut fonctionner sur cet ordinateur. Utilisez plutôt un serveur.",
   "onboarding_chat_step_title": "Choisissez un modèle de génération de texte",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "Choisissez d’abord un modèle de chat",
   "onboarding_hugging_face_search_hide_button": "Masquer la recherche Hugging Face",
   "onboarding_hugging_face_search_show_button": "Absent de la liste ? Rechercher sur Hugging Face",
-  "onboarding_image_edit_step_body": "Modifie vos images. Exécutez-en un sur cet ordinateur ou utilisez-en un depuis un serveur.",
-  "onboarding_image_edit_step_no_local_empty": "Les modèles de modification d’images ne peuvent pas fonctionner sur cet ordinateur. Utilisez plutôt un serveur.",
-  "onboarding_image_edit_step_title": "Choisissez un modèle de modification d’images",
-  "onboarding_image_step_body": "Crée vos images. Exécutez-en un sur cet ordinateur ou utilisez-en un depuis un serveur.",
-  "onboarding_image_step_no_local_empty": "Les modèles d’image ne peuvent pas fonctionner sur cet ordinateur. Utilisez plutôt un serveur.",
-  "onboarding_image_step_title": "Choisissez un modèle de génération d’images",
   "onboarding_model_list_aria": "Modèles pour cet ordinateur",
   "onboarding_model_list_delete_aria": "Supprimer {name}",
   "onboarding_model_list_hugging_face_label": "Depuis Hugging Face",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "Sur cet ordinateur",
   "onboarding_model_step_local_title": "Sur cet ordinateur",
   "onboarding_model_step_optional_label": "Facultatif",
-  "onboarding_model_step_skip_button": "Ignorer",
   "onboarding_model_step_skip_finish_button": "Ignorer et terminer",
   "onboarding_offline_start_body": "Démarrez-le avec cette commande :",
   "onboarding_offline_title": "Backend local indisponible",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "Connecter un autre serveur",
   "onboarding_server_path_connect_button": "Connecter un serveur",
   "onboarding_server_path_empty": "Aucun serveur connecté pour l’instant.",
-  "onboarding_video_step_body": "Crée de courts clips vidéo. Exécutez-en un sur cet ordinateur ou utilisez-en un depuis un serveur.",
-  "onboarding_video_step_no_local_empty": "Les modèles vidéo ne peuvent pas fonctionner sur cet ordinateur. Utilisez plutôt un serveur.",
-  "onboarding_video_step_title": "Choisissez un modèle de génération de vidéos",
   "onboarding_welcome_body": "Un carnet de recherche privé qui fonctionne entièrement sur votre propre machine. Vos documents, vos clés de modèle, sans cloud, sans compte.",
   "onboarding_welcome_start_button": "Commencer la configuration",
   "onboarding_welcome_title": "L’<accent>alternative à NotebookLM</accent> air-gapped et open source",
diff --git a/surfsense_local/frontend/translations/hi.json b/surfsense_local/frontend/translations/hi.json
index 19ae883ad7..32f27319d5 100644
--- a/surfsense_local/frontend/translations/hi.json
+++ b/surfsense_local/frontend/translations/hi.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Office सपोर्ट",
   "office_support_settings_turn_on_button": "चालू करें",
   "office_support_settings_using_yours_status": "{path} पर आपका अपना इंस्टॉलेशन",
-  "onboarding_audio_step_body": "आपके लिए पॉडकास्ट बनाता है। इसे इस कंप्यूटर पर चलाएं, या किसी सर्वर से इस्तेमाल करें।",
-  "onboarding_audio_step_no_local_empty": "ऑडियो मॉडल इस कंप्यूटर पर नहीं चल सकते। इसके बजाय कोई सर्वर इस्तेमाल करें।",
-  "onboarding_audio_step_title": "ऑडियो मॉडल चुनें",
   "onboarding_chat_step_body": "चैट में आपको जवाब देता है। इसे इस कंप्यूटर पर चलाएं ताकि आपकी चैट निजी रहें, या किसी सर्वर से इस्तेमाल करें।",
   "onboarding_chat_step_no_local_empty": "इस कंप्यूटर पर कोई भी जांचा गया मॉडल नहीं चल सकता। इसके बजाय कोई सर्वर इस्तेमाल करें।",
   "onboarding_chat_step_title": "टेक्स्ट जनरेशन मॉडल चुनें",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "पहले चैट मॉडल चुनें",
   "onboarding_hugging_face_search_hide_button": "Hugging Face खोज छिपाएं",
   "onboarding_hugging_face_search_show_button": "सूची में नहीं है? Hugging Face पर खोजें",
-  "onboarding_image_edit_step_body": "आपके लिए इमेज एडिट करता है। इसे इस कंप्यूटर पर चलाएं, या किसी सर्वर से इस्तेमाल करें।",
-  "onboarding_image_edit_step_no_local_empty": "इमेज एडिटिंग मॉडल इस कंप्यूटर पर नहीं चल सकते। इसके बजाय कोई सर्वर इस्तेमाल करें।",
-  "onboarding_image_edit_step_title": "इमेज एडिटिंग मॉडल चुनें",
-  "onboarding_image_step_body": "आपके लिए इमेज बनाता है। इसे इस कंप्यूटर पर चलाएं, या किसी सर्वर से इस्तेमाल करें।",
-  "onboarding_image_step_no_local_empty": "इमेज मॉडल इस कंप्यूटर पर नहीं चल सकते। इसके बजाय कोई सर्वर इस्तेमाल करें।",
-  "onboarding_image_step_title": "इमेज जनरेशन मॉडल चुनें",
   "onboarding_model_list_aria": "इस कंप्यूटर के लिए मॉडल",
   "onboarding_model_list_delete_aria": "{name} हटाएं",
   "onboarding_model_list_hugging_face_label": "Hugging Face से",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "इस कंप्यूटर पर",
   "onboarding_model_step_local_title": "इस कंप्यूटर पर",
   "onboarding_model_step_optional_label": "वैकल्पिक",
-  "onboarding_model_step_skip_button": "छोड़ें",
   "onboarding_model_step_skip_finish_button": "छोड़ें और समाप्त करें",
   "onboarding_offline_start_body": "इसे इस कमांड से शुरू करें:",
   "onboarding_offline_title": "लोकल बैकएंड उपलब्ध नहीं है",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "दूसरा सर्वर कनेक्ट करें",
   "onboarding_server_path_connect_button": "सर्वर कनेक्ट करें",
   "onboarding_server_path_empty": "अभी तक कोई सर्वर कनेक्ट नहीं है।",
-  "onboarding_video_step_body": "छोटे वीडियो क्लिप बनाता है। इसे इस कंप्यूटर पर चलाएं, या किसी सर्वर से इस्तेमाल करें।",
-  "onboarding_video_step_no_local_empty": "वीडियो मॉडल इस कंप्यूटर पर नहीं चल सकते। इसके बजाय कोई सर्वर इस्तेमाल करें।",
-  "onboarding_video_step_title": "वीडियो जनरेशन मॉडल चुनें",
   "onboarding_welcome_body": "एक निजी रिसर्च नोटबुक जो पूरी तरह आपकी अपनी मशीन पर चलती है। आपके दस्तावेज़, आपकी मॉडल कुंजियां, न कोई क्लाउड, न कोई खाता।",
   "onboarding_welcome_start_button": "सेटअप शुरू करें",
   "onboarding_welcome_title": "एयर-गैप्ड, ओपन सोर्स <accent>NotebookLM विकल्प</accent>",
diff --git a/surfsense_local/frontend/translations/ja.json b/surfsense_local/frontend/translations/ja.json
index f4791332c0..b6c53b9d8e 100644
--- a/surfsense_local/frontend/translations/ja.json
+++ b/surfsense_local/frontend/translations/ja.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Officeサポート",
   "office_support_settings_turn_on_button": "オンにする",
   "office_support_settings_using_yours_status": "{path}にあるご自身のインストール",
-  "onboarding_audio_step_body": "ポッドキャストを作成します。このコンピューターで実行するか、サーバー上のモデルを使います。",
-  "onboarding_audio_step_no_local_empty": "このコンピューターでは音声モデルを実行できません。代わりにサーバーを使ってください。",
-  "onboarding_audio_step_title": "音声モデルを選択",
   "onboarding_chat_step_body": "チャットで回答します。このコンピューターで実行すればチャットはプライベートなまま保たれます。サーバー上のモデルを使うこともできます。",
   "onboarding_chat_step_no_local_empty": "このコンピューターで実行できる検証済みのモデルはありません。代わりにサーバーを使ってください。",
   "onboarding_chat_step_title": "テキスト生成モデルを選択",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "先にチャットモデルを選んでください",
   "onboarding_hugging_face_search_hide_button": "Hugging Faceの検索を閉じる",
   "onboarding_hugging_face_search_show_button": "見つからない場合はHugging Faceを検索",
-  "onboarding_image_edit_step_body": "画像を編集します。このコンピューターで実行するか、サーバー上のモデルを使います。",
-  "onboarding_image_edit_step_no_local_empty": "このコンピューターでは画像編集モデルを実行できません。代わりにサーバーを使ってください。",
-  "onboarding_image_edit_step_title": "画像編集モデルを選択",
-  "onboarding_image_step_body": "画像を作成します。このコンピューターで実行するか、サーバー上のモデルを使います。",
-  "onboarding_image_step_no_local_empty": "このコンピューターでは画像モデルを実行できません。代わりにサーバーを使ってください。",
-  "onboarding_image_step_title": "画像生成モデルを選択",
   "onboarding_model_list_aria": "このコンピューター向けのモデル",
   "onboarding_model_list_delete_aria": "{name}を削除",
   "onboarding_model_list_hugging_face_label": "Hugging Faceから",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "このコンピューター上",
   "onboarding_model_step_local_title": "このコンピューター上",
   "onboarding_model_step_optional_label": "任意",
-  "onboarding_model_step_skip_button": "スキップ",
   "onboarding_model_step_skip_finish_button": "スキップして完了",
   "onboarding_offline_start_body": "次のコマンドで起動してください：",
   "onboarding_offline_title": "ローカルバックエンドを利用できません",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "別のサーバーを接続",
   "onboarding_server_path_connect_button": "サーバーを接続",
   "onboarding_server_path_empty": "接続済みのサーバーはまだありません。",
-  "onboarding_video_step_body": "短い動画クリップを作成します。このコンピューターで実行するか、サーバー上のモデルを使います。",
-  "onboarding_video_step_no_local_empty": "このコンピューターでは動画モデルを実行できません。代わりにサーバーを使ってください。",
-  "onboarding_video_step_title": "動画生成モデルを選択",
   "onboarding_welcome_body": "完全に自分のマシン上で動作するプライベートなリサーチノートブックです。ドキュメントもモデルのキーも手元に。クラウドもアカウントも不要です。",
   "onboarding_welcome_start_button": "セットアップを始める",
   "onboarding_welcome_title": "エアギャップ対応、オープンソースの<accent>NotebookLMの代替</accent>",
diff --git a/surfsense_local/frontend/translations/ko.json b/surfsense_local/frontend/translations/ko.json
index 0683cb9c8d..d9e537aeab 100644
--- a/surfsense_local/frontend/translations/ko.json
+++ b/surfsense_local/frontend/translations/ko.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Office 지원",
   "office_support_settings_turn_on_button": "켜기",
   "office_support_settings_using_yours_status": "{path}에 직접 설치한 버전",
-  "onboarding_audio_step_body": "팟캐스트를 만들어 줍니다. 이 컴퓨터에서 실행하거나, 서버의 모델을 사용하세요.",
-  "onboarding_audio_step_no_local_empty": "이 컴퓨터에서는 음성 모델을 실행할 수 없습니다. 대신 서버를 사용하세요.",
-  "onboarding_audio_step_title": "음성 모델 선택",
   "onboarding_chat_step_body": "채팅에서 답변합니다. 이 컴퓨터에서 실행하면 채팅이 비공개로 유지되며, 서버의 모델을 사용할 수도 있습니다.",
   "onboarding_chat_step_no_local_empty": "이 컴퓨터에서 실행할 수 있는 테스트된 모델이 없습니다. 대신 서버를 사용하세요.",
   "onboarding_chat_step_title": "텍스트 생성 모델 선택",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "먼저 채팅 모델을 선택하세요",
   "onboarding_hugging_face_search_hide_button": "Hugging Face 검색 숨기기",
   "onboarding_hugging_face_search_show_button": "목록에 없나요? Hugging Face 검색",
-  "onboarding_image_edit_step_body": "이미지를 편집해 줍니다. 이 컴퓨터에서 실행하거나, 서버의 모델을 사용하세요.",
-  "onboarding_image_edit_step_no_local_empty": "이 컴퓨터에서는 이미지 편집 모델을 실행할 수 없습니다. 대신 서버를 사용하세요.",
-  "onboarding_image_edit_step_title": "이미지 편집 모델 선택",
-  "onboarding_image_step_body": "이미지를 만들어 줍니다. 이 컴퓨터에서 실행하거나, 서버의 모델을 사용하세요.",
-  "onboarding_image_step_no_local_empty": "이 컴퓨터에서는 이미지 모델을 실행할 수 없습니다. 대신 서버를 사용하세요.",
-  "onboarding_image_step_title": "이미지 생성 모델 선택",
   "onboarding_model_list_aria": "이 컴퓨터용 모델",
   "onboarding_model_list_delete_aria": "{name} 삭제",
   "onboarding_model_list_hugging_face_label": "Hugging Face에서",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "이 컴퓨터에서",
   "onboarding_model_step_local_title": "이 컴퓨터에서",
   "onboarding_model_step_optional_label": "선택 사항",
-  "onboarding_model_step_skip_button": "건너뛰기",
   "onboarding_model_step_skip_finish_button": "건너뛰고 완료",
   "onboarding_offline_start_body": "다음 명령으로 시작하세요:",
   "onboarding_offline_title": "로컬 백엔드를 사용할 수 없습니다",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "다른 서버 연결",
   "onboarding_server_path_connect_button": "서버 연결",
   "onboarding_server_path_empty": "아직 연결된 서버가 없습니다.",
-  "onboarding_video_step_body": "짧은 동영상 클립을 만들어 줍니다. 이 컴퓨터에서 실행하거나, 서버의 모델을 사용하세요.",
-  "onboarding_video_step_no_local_empty": "이 컴퓨터에서는 동영상 모델을 실행할 수 없습니다. 대신 서버를 사용하세요.",
-  "onboarding_video_step_title": "동영상 생성 모델 선택",
   "onboarding_welcome_body": "전적으로 내 컴퓨터에서 실행되는 비공개 리서치 노트북입니다. 문서도 모델 키도 내 것이며, 클라우드도 계정도 없습니다.",
   "onboarding_welcome_start_button": "설정 시작",
   "onboarding_welcome_title": "에어갭 지원 오픈소스 <accent>NotebookLM 대안</accent>",
diff --git a/surfsense_local/frontend/translations/pt-BR.json b/surfsense_local/frontend/translations/pt-BR.json
index 6c45d1e466..88b1ed58e9 100644
--- a/surfsense_local/frontend/translations/pt-BR.json
+++ b/surfsense_local/frontend/translations/pt-BR.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Suporte ao Office",
   "office_support_settings_turn_on_button": "Ativar",
   "office_support_settings_using_yours_status": "Sua própria instalação em {path}",
-  "onboarding_audio_step_body": "Cria podcasts para você. Execute um neste computador ou use um de um servidor.",
-  "onboarding_audio_step_no_local_empty": "Modelos de áudio não podem ser executados neste computador. Em vez disso, use um servidor.",
-  "onboarding_audio_step_title": "Escolha um modelo de áudio",
   "onboarding_chat_step_body": "Responde a você no chat. Execute um neste computador para que seus chats fiquem privados, ou use um de um servidor.",
   "onboarding_chat_step_no_local_empty": "Nenhum modelo testado pode ser executado neste computador. Em vez disso, use um servidor.",
   "onboarding_chat_step_title": "Escolha um modelo de geração de texto",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "Escolha primeiro um modelo de chat",
   "onboarding_hugging_face_search_hide_button": "Ocultar a busca no Hugging Face",
   "onboarding_hugging_face_search_show_button": "Não está na lista? Busque no Hugging Face",
-  "onboarding_image_edit_step_body": "Edita imagens para você. Execute um neste computador ou use um de um servidor.",
-  "onboarding_image_edit_step_no_local_empty": "Modelos de edição de imagem não podem ser executados neste computador. Em vez disso, use um servidor.",
-  "onboarding_image_edit_step_title": "Escolha um modelo de edição de imagem",
-  "onboarding_image_step_body": "Cria imagens para você. Execute um neste computador ou use um de um servidor.",
-  "onboarding_image_step_no_local_empty": "Modelos de imagem não podem ser executados neste computador. Em vez disso, use um servidor.",
-  "onboarding_image_step_title": "Escolha um modelo de geração de imagem",
   "onboarding_model_list_aria": "Modelos para este computador",
   "onboarding_model_list_delete_aria": "Excluir {name}",
   "onboarding_model_list_hugging_face_label": "Do Hugging Face",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "Neste computador",
   "onboarding_model_step_local_title": "Neste computador",
   "onboarding_model_step_optional_label": "Opcional",
-  "onboarding_model_step_skip_button": "Pular",
   "onboarding_model_step_skip_finish_button": "Pular e concluir",
   "onboarding_offline_start_body": "Inicie-o com este comando:",
   "onboarding_offline_title": "Backend local indisponível",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "Conectar outro servidor",
   "onboarding_server_path_connect_button": "Conectar um servidor",
   "onboarding_server_path_empty": "Nenhum servidor conectado ainda.",
-  "onboarding_video_step_body": "Cria clipes de vídeo curtos. Execute um neste computador ou use um de um servidor.",
-  "onboarding_video_step_no_local_empty": "Modelos de vídeo não podem ser executados neste computador. Em vez disso, use um servidor.",
-  "onboarding_video_step_title": "Escolha um modelo de geração de vídeo",
   "onboarding_welcome_body": "Um caderno de pesquisa privado que roda inteiramente na sua própria máquina. Seus documentos, suas chaves de modelo, sem nuvem, sem conta.",
   "onboarding_welcome_start_button": "Começar a configurar",
   "onboarding_welcome_title": "<accent>Alternativa ao NotebookLM</accent> air-gapped e de código aberto",
diff --git a/surfsense_local/frontend/translations/ru.json b/surfsense_local/frontend/translations/ru.json
index 73c4e55ba7..43c6aa365b 100644
--- a/surfsense_local/frontend/translations/ru.json
+++ b/surfsense_local/frontend/translations/ru.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Поддержка Office",
   "office_support_settings_turn_on_button": "Включить",
   "office_support_settings_using_yours_status": "Ваша установка в {path}",
-  "onboarding_audio_step_body": "Создаёт для вас подкасты. Запустите её на этом компьютере или используйте модель с сервера.",
-  "onboarding_audio_step_no_local_empty": "Модели для аудио не запускаются на этом компьютере. Используйте вместо них сервер.",
-  "onboarding_audio_step_title": "Выберите модель для аудио",
   "onboarding_chat_step_body": "Отвечает вам в чате. Запустите её на этом компьютере, чтобы чаты оставались приватными, или используйте модель с сервера.",
   "onboarding_chat_step_no_local_empty": "Ни одна проверенная модель не запускается на этом компьютере. Используйте вместо неё сервер.",
   "onboarding_chat_step_title": "Выберите модель генерации текста",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "Сначала выберите модель для чата",
   "onboarding_hugging_face_search_hide_button": "Скрыть поиск по Hugging Face",
   "onboarding_hugging_face_search_show_button": "Нет в списке? Поищите на Hugging Face",
-  "onboarding_image_edit_step_body": "Редактирует для вас изображения. Запустите её на этом компьютере или используйте модель с сервера.",
-  "onboarding_image_edit_step_no_local_empty": "Модели для редактирования изображений не запускаются на этом компьютере. Используйте вместо них сервер.",
-  "onboarding_image_edit_step_title": "Выберите модель редактирования изображений",
-  "onboarding_image_step_body": "Создаёт для вас изображения. Запустите её на этом компьютере или используйте модель с сервера.",
-  "onboarding_image_step_no_local_empty": "Модели для изображений не запускаются на этом компьютере. Используйте вместо них сервер.",
-  "onboarding_image_step_title": "Выберите модель генерации изображений",
   "onboarding_model_list_aria": "Модели для этого компьютера",
   "onboarding_model_list_delete_aria": "Удалить {name}",
   "onboarding_model_list_hugging_face_label": "С Hugging Face",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "На этом компьютере",
   "onboarding_model_step_local_title": "На этом компьютере",
   "onboarding_model_step_optional_label": "Необязательно",
-  "onboarding_model_step_skip_button": "Пропустить",
   "onboarding_model_step_skip_finish_button": "Пропустить и завершить",
   "onboarding_offline_start_body": "Запустите его этой командой:",
   "onboarding_offline_title": "Локальный бэкенд недоступен",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "Подключить другой сервер",
   "onboarding_server_path_connect_button": "Подключить сервер",
   "onboarding_server_path_empty": "Пока не подключено ни одного сервера.",
-  "onboarding_video_step_body": "Создаёт короткие видеоклипы. Запустите её на этом компьютере или используйте модель с сервера.",
-  "onboarding_video_step_no_local_empty": "Модели для видео не запускаются на этом компьютере. Используйте вместо них сервер.",
-  "onboarding_video_step_title": "Выберите модель генерации видео",
   "onboarding_welcome_body": "Личный исследовательский блокнот, который работает целиком на вашей машине. Ваши документы, ваши ключи моделей, без облака и без аккаунта.",
   "onboarding_welcome_start_button": "Начать настройку",
   "onboarding_welcome_title": "<accent>Аналог NotebookLM</accent> с открытым исходным кодом, изолированный от сети",
diff --git a/surfsense_local/frontend/translations/zh-CN.json b/surfsense_local/frontend/translations/zh-CN.json
index b7b5405605..9dd3071660 100644
--- a/surfsense_local/frontend/translations/zh-CN.json
+++ b/surfsense_local/frontend/translations/zh-CN.json
@@ -828,9 +828,6 @@
   "office_support_settings_title": "Office 支持",
   "office_support_settings_turn_on_button": "开启",
   "office_support_settings_using_yours_status": "你自己安装在 {path} 的版本",
-  "onboarding_audio_step_body": "为你制作播客。可以在这台电脑上运行一个，也可以使用服务器上的。",
-  "onboarding_audio_step_no_local_empty": "这台电脑无法运行语音模型。请改用服务器。",
-  "onboarding_audio_step_title": "选择语音模型",
   "onboarding_chat_step_body": "在对话中回答你。在这台电脑上运行可以让你的对话保持私密，也可以使用服务器上的。",
   "onboarding_chat_step_no_local_empty": "这台电脑无法运行任何经过测试的模型。请改用服务器。",
   "onboarding_chat_step_title": "选择文本生成模型",
@@ -844,12 +841,6 @@
   "onboarding_finish_no_chat_model_error": "请先选择对话模型",
   "onboarding_hugging_face_search_hide_button": "隐藏 Hugging Face 搜索",
   "onboarding_hugging_face_search_show_button": "没有列出？搜索 Hugging Face",
-  "onboarding_image_edit_step_body": "为你编辑图像。可以在这台电脑上运行一个，也可以使用服务器上的。",
-  "onboarding_image_edit_step_no_local_empty": "这台电脑无法运行图像编辑模型。请改用服务器。",
-  "onboarding_image_edit_step_title": "选择图像编辑模型",
-  "onboarding_image_step_body": "为你生成图像。可以在这台电脑上运行一个，也可以使用服务器上的。",
-  "onboarding_image_step_no_local_empty": "这台电脑无法运行图像模型。请改用服务器。",
-  "onboarding_image_step_title": "选择图像生成模型",
   "onboarding_model_list_aria": "适用于这台电脑的模型",
   "onboarding_model_list_delete_aria": "删除 {name}",
   "onboarding_model_list_hugging_face_label": "来自 Hugging Face",
@@ -868,7 +859,6 @@
   "onboarding_model_step_local_aria": "这台电脑上",
   "onboarding_model_step_local_title": "这台电脑上",
   "onboarding_model_step_optional_label": "可选",
-  "onboarding_model_step_skip_button": "跳过",
   "onboarding_model_step_skip_finish_button": "跳过并完成",
   "onboarding_offline_start_body": "用这条命令启动它：",
   "onboarding_offline_title": "本地后端不可用",
@@ -895,9 +885,6 @@
   "onboarding_server_path_connect_another_button": "连接另一台服务器",
   "onboarding_server_path_connect_button": "连接服务器",
   "onboarding_server_path_empty": "还没有连接任何服务器。",
-  "onboarding_video_step_body": "生成短视频片段。可以在这台电脑上运行一个，也可以使用服务器上的。",
-  "onboarding_video_step_no_local_empty": "这台电脑无法运行视频模型。请改用服务器。",
-  "onboarding_video_step_title": "选择视频生成模型",
   "onboarding_welcome_body": "完全在你自己的机器上运行的私密研究笔记本。你的文档，你的模型密钥，没有云端，不用账户。",
   "onboarding_welcome_start_button": "开始设置",
   "onboarding_welcome_title": "气隙隔离、开源的 <accent>NotebookLM 替代品</accent>",
diff --git a/www/content/docs/v2/models/api-providers.mdx b/www/content/docs/v2/models/api-providers.mdx
index ac21487d17..4e78ad6ce1 100644
--- a/www/content/docs/v2/models/api-providers.mdx
+++ b/www/content/docs/v2/models/api-providers.mdx
@@ -12,7 +12,7 @@ You can use models from a hosted provider, such as OpenAI, Anthropic, OpenRouter
 <Steps>
 <Step>
 
-Open **Settings** and the model section you want, such as **Text gen**. Click **Add model**, then **Use a server**. In onboarding, click **Connect** under **Use a server**.
+Open **Settings** and the model section you want, such as **Text gen**. Click **Add model**, then **Use a server**. In onboarding, the text generation step has **Use a server**: click **Connect**.
 
 </Step>
 <Step>
diff --git a/www/content/docs/v2/models/index.mdx b/www/content/docs/v2/models/index.mdx
index 0437b9a13c..008eebd5bf 100644
--- a/www/content/docs/v2/models/index.mdx
+++ b/www/content/docs/v2/models/index.mdx
@@ -47,7 +47,7 @@ Open **Settings** and look under **Models**. Each section shows the model **In u
 - Each server you connected has its own group, with **Edit** and **Disconnect**. Open the group to see its models.
 - **Add model** opens a page with **On this computer**, **Use a server**, and for text generation, **Use your ChatGPT plan** and **Use your Claude plan**.
 
-When you first open SurfSense, onboarding walks through the same choices, one kind of model per step. Everything except the text generation model can be skipped, and you can change all of it later in Settings, except the embedding model.
+When you first open SurfSense, onboarding asks for two models: a text generation model, which is required, and an embedding model, which can be skipped. You choose the others in Settings when you want them. You can change all of them later, except the embedding model.
 
 ## What fits your computer
 
diff --git a/www/content/docs/v2/models/other-local-servers.mdx b/www/content/docs/v2/models/other-local-servers.mdx
index fd7c4d496d..a90c64ebd8 100644
--- a/www/content/docs/v2/models/other-local-servers.mdx
+++ b/www/content/docs/v2/models/other-local-servers.mdx
@@ -26,7 +26,7 @@ SurfSense sends no settings specific to one server, so a strict server accepts i
 <Steps>
 <Step>
 
-Open **Settings** and the model section you want, such as **Text gen**. Click **Add model**, then **Use a server**. In onboarding, click **Connect** under **Use a server**.
+Open **Settings** and the model section you want, such as **Text gen**. Click **Add model**, then **Use a server**. In onboarding, the text generation step has **Use a server**: click **Connect**.
 
 </Step>
 <Step>
diff --git a/www/content/docs/v2/models/other-models.mdx b/www/content/docs/v2/models/other-models.mdx
index 7602a074f8..f18e8fbcdc 100644
--- a/www/content/docs/v2/models/other-models.mdx
+++ b/www/content/docs/v2/models/other-models.mdx
@@ -30,7 +30,7 @@ Besides the chat model, SurfSense uses a model for each of these jobs. All are o
 
 **Settings › Image edit** holds the model that changes an image Studio made, from an instruction such as "use a darker colour scheme". Each edit becomes the image's next version. See [Studio](/docs/v2/studio).
 
-- **On this computer**, FLUX.2 klein 4B both draws and edits from one download, so if you use it for images, choose it here too. In onboarding, the editing step lists it first when you chose it for images. Editing on a processor without a graphics card is slow: expect many minutes per edit.
+- **On this computer**, FLUX.2 klein 4B both draws and edits from one download, so if you use it for images, choose it here too. Editing on a processor without a graphics card is slow: expect many minutes per edit.
 - **From a server**, use a model that takes a picture in and gives one back. The server must implement `/images/edits`.
 
 Without an image editing model, Studio's edit box says so and links to this section.

From 4c04519e60a5f9ce8753e57611d171c3a4a1993b Mon Sep 17 00:00:00 2001
From: Anish Sarkar <104695310+AnishSarkar22@users.noreply.github.com>
Date: Wed, 4 Nov 2026 17:13:52 +0530
Subject: [PATCH 12/12] fix(local): address review findings on startup, motion
 and site logos

A saved mode pick now updates what the next launch's startup frame reads,
so it no longer offers mentions for a chat that opens in Basic. A Reduce
motion choice storage refuses still applies until the page reloads. A site
logo that failed no longer leaves the next site on its letter. The test of
a short startup wait freezes the clock, so a slow machine cannot fail it.
---
 .../src/app/app-bootstrap-startup.test.tsx    | 13 +++++-
 .../features/chat/modes/mode-picker.test.tsx  | 42 +++++++++++++++++++
 .../features/chat/modes/next-new-chat-mode.ts |  4 +-
 .../src/features/chat/modes/save-mode-pick.ts |  9 +++-
 .../motion-preference.test.ts                 | 30 ++++++++++++-
 .../motion-preference/motion-preference.ts    |  7 +++-
 .../sources/web/search-site-logo.test.tsx     | 20 +++++++++
 .../features/sources/web/search-site-logo.tsx |  2 +
 8 files changed, 120 insertions(+), 7 deletions(-)
 create mode 100644 surfsense_local/frontend/src/features/sources/web/search-site-logo.test.tsx

diff --git a/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx b/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
index f7f148c2f6..1dd113af1a 100644
--- a/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
+++ b/surfsense_local/frontend/src/app/app-bootstrap-startup.test.tsx
@@ -6,7 +6,15 @@ import {
   within,
 } from "@testing-library/react"
 import userEvent from "@testing-library/user-event"
-import { afterEach, beforeEach, describe, expect, it, vi } from "vitest"
+import {
+  afterEach,
+  beforeEach,
+  describe,
+  expect,
+  it,
+  onTestFinished,
+  vi,
+} from "vitest"
 
 import { TooltipProvider } from "@/components/ui/tooltip"
 import { writeNextNewChatMode } from "@/features/chat/modes/next-new-chat-mode-store"
@@ -415,6 +423,9 @@ describe("startup's reveal", () => {
   }, 15_000)
 
   it("draws them at once after a wait too short to see", async () => {
+    // The wait is measured on this clock, so a slow machine cannot stretch it.
+    const clock = vi.spyOn(Date, "now").mockReturnValue(Date.now())
+    onTestFinished(() => clock.mockRestore())
     onDatabase("a")
     rememberDashboard()
     completedInstall()
diff --git a/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx b/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
index 992b41d607..4f3df401f4 100644
--- a/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
+++ b/surfsense_local/frontend/src/features/chat/modes/mode-picker.test.tsx
@@ -5,11 +5,13 @@ import { AssistantRuntimeProvider, useLocalRuntime } from "@assistant-ui/react"
 
 import type { ChatMode, ChatModes } from "@/features/models/capability/api"
 import type { ModelSelection } from "@/features/models/selection/api"
+import { readStartupLayout } from "@/features/startup/startup-layout"
 import { render } from "@/test-utils"
 
 import { ChatComposer } from "../chat-composer"
 import { useNewChatMode } from "./new-chat-mode"
 import { NewChatModeProvider } from "./new-chat-mode-provider"
+import { rememberNextNewChatMode } from "./next-new-chat-mode"
 
 const PASSED: ChatModes = {
   agentic_allowed: true,
@@ -165,6 +167,46 @@ describe("the composer's mode switch", () => {
     expect(nextChat()).toBe("basic")
   })
 
+  it("tells the next launch's startup frame the mode a saved pick made the default", async () => {
+    rememberNextNewChatMode(KIMI)
+    const user = userEvent.setup()
+    renderComposer({ model: KIMI })
+    vi.mocked(fetch).mockImplementation(async () =>
+      Response.json({ ...PASSED, default_mode: "basic" })
+    )
+    expect(readStartupLayout().mentions).toBe(true)
+
+    await user.click(trigger())
+    await user.click(
+      await screen.findByRole("menuitemradio", { name: /^Basic \(Q&A\)/ })
+    )
+
+    await waitFor(() => expect(readStartupLayout().mentions).toBe(false))
+  })
+
+  it("leaves the startup frame's mode alone when a pick could not be saved", async () => {
+    rememberNextNewChatMode(KIMI)
+    const user = userEvent.setup()
+    renderComposer({ model: KIMI })
+    const fetchMock = vi.mocked(fetch)
+    fetchMock.mockImplementation(
+      async () => new Response("Internal Server Error", { status: 500 })
+    )
+
+    await user.click(trigger())
+    await user.click(
+      await screen.findByRole("menuitemradio", { name: /^Basic \(Q&A\)/ })
+    )
+
+    await waitFor(() =>
+      expect(fetchMock).toHaveBeenCalledWith(
+        expect.stringContaining("/llm/selection/text_gen/mode"),
+        expect.anything()
+      )
+    )
+    expect(readStartupLayout().mentions).toBe(true)
+  })
+
   it("starts an untested model in Basic, and warns only in the menu", async () => {
     const user = userEvent.setup()
     renderComposer({ model: SONNET_99 })
diff --git a/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts
index 39fa3c043e..60d9ef0ec4 100644
--- a/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts
+++ b/surfsense_local/frontend/src/features/chat/modes/next-new-chat-mode.ts
@@ -5,8 +5,8 @@ import { writeNextNewChatMode } from "./next-new-chat-mode-store"
 
 /**
  * Remembers the mode a new chat on this model opens in after a restart: its
- * default, since a mode picked this session is not kept. The startup frame
- * reads it to offer plugin mentions before the model has loaded.
+ * default, which a saved pick becomes. The startup frame reads it to offer
+ * plugin mentions before the model has loaded.
  */
 export function rememberNextNewChatMode(model: ModelSelection | null): void {
   writeNextNewChatMode(resolveNewChatChoice(model, {})?.mode ?? "basic")
diff --git a/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts b/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
index 706184699d..ad78e31a9c 100644
--- a/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
+++ b/surfsense_local/frontend/src/features/chat/modes/save-mode-pick.ts
@@ -9,6 +9,7 @@ import { intl } from "@/i18n/intl"
 import { requestJson } from "@/lib/api"
 
 import { usePickNewChatMode } from "./new-chat-mode"
+import { writeNextNewChatMode } from "./next-new-chat-mode-store"
 
 function saveModePick(
   model: ModelSelection,
@@ -38,14 +39,18 @@ export function useSaveModePick() {
     (model: ModelSelection, mode: ChatMode) => {
       pick(model, mode)
       saveModePick(model, mode).then(
-        (modes) =>
+        (modes) => {
+          // For the startup frame: the dashboard writes this when the model
+          // changes, and a pick does not change it.
+          writeNextNewChatMode(modes.default_mode)
           client.setQueryData<ModelSelection | null>(
             selectionQueryKey("text_gen"),
             (current) =>
               current?.capability && modelKey(current) === modelKey(model)
                 ? { ...current, capability: { ...current.capability, modes } }
                 : current
-          ),
+          )
+        },
         () =>
           toast.error(
             intl.formatMessage({
diff --git a/surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts b/surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts
index 7215b732ab..576af4a8dd 100644
--- a/surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts
+++ b/surfsense_local/frontend/src/features/motion-preference/motion-preference.test.ts
@@ -1,4 +1,12 @@
-import { afterEach, beforeEach, describe, expect, it, vi } from "vitest"
+import {
+  afterEach,
+  beforeEach,
+  describe,
+  expect,
+  it,
+  onTestFinished,
+  vi,
+} from "vitest"
 
 import indexHtml from "../../../index.html?raw"
 
@@ -6,6 +14,7 @@ import {
   followMotionPreference,
   MOTION_PREFERENCE_KEY,
   motionIsReduced,
+  readMotionPreference,
   setMotionPreference,
 } from "./motion-preference"
 
@@ -98,6 +107,25 @@ describe("motion preference", () => {
     expect(localStorage.getItem(MOTION_PREFERENCE_KEY)).toBe("off")
   })
 
+  it("still applies a choice storage refuses to keep", () => {
+    osAsksToReduce(false)
+    const write = vi
+      .spyOn(Storage.prototype, "setItem")
+      .mockImplementation(() => {
+        throw new DOMException("blocked", "SecurityError")
+      })
+    onTestFinished(() => {
+      write.mockRestore()
+      setMotionPreference("system")
+    })
+
+    setMotionPreference("on")
+
+    expect(root.classList.contains(CLASS)).toBe(true)
+    expect(motionIsReduced()).toBe(true)
+    expect(readMotionPreference()).toBe("on")
+  })
+
   it("follows the OS while it matches the system, and ignores it otherwise", () => {
     const os = osAsksToReduce(false)
     followMotionPreference()
diff --git a/surfsense_local/frontend/src/features/motion-preference/motion-preference.ts b/surfsense_local/frontend/src/features/motion-preference/motion-preference.ts
index f6dd1d7d87..a84289cc00 100644
--- a/surfsense_local/frontend/src/features/motion-preference/motion-preference.ts
+++ b/surfsense_local/frontend/src/features/motion-preference/motion-preference.ts
@@ -6,7 +6,11 @@ export const MOTION_PREFERENCE_KEY = "surfsense:reduce-motion:v1"
 const REDUCED_MOTION_QUERY = "(prefers-reduced-motion: reduce)"
 const ROOT_CLASS = "reduce-motion"
 
+// A choice storage refused: it holds until the page reloads.
+let unsaved: MotionPreference | null = null
+
 export function readMotionPreference(): MotionPreference {
+  if (unsaved) return unsaved
   try {
     const saved = localStorage.getItem(MOTION_PREFERENCE_KEY)
     return saved === "on" || saved === "off" ? saved : "system"
@@ -32,8 +36,9 @@ function applyToRoot(): void {
 export function setMotionPreference(preference: MotionPreference): void {
   try {
     localStorage.setItem(MOTION_PREFERENCE_KEY, preference)
+    unsaved = null
   } catch {
-    // Without storage, the choice cannot be kept.
+    unsaved = preference
   }
   applyToRoot()
 }
diff --git a/surfsense_local/frontend/src/features/sources/web/search-site-logo.test.tsx b/surfsense_local/frontend/src/features/sources/web/search-site-logo.test.tsx
new file mode 100644
index 0000000000..677429b93b
--- /dev/null
+++ b/surfsense_local/frontend/src/features/sources/web/search-site-logo.test.tsx
@@ -0,0 +1,20 @@
+import { cleanup, fireEvent } from "@testing-library/react"
+import { afterEach, describe, expect, it } from "vitest"
+
+import { render } from "@/test-utils"
+
+import { SearchSiteLogo } from "./search-site-logo"
+
+afterEach(cleanup)
+
+describe("SearchSiteLogo", () => {
+  it("tries the next site's logo after the last one failed", () => {
+    const { container, rerender } = render(<SearchSiteLogo site="reddit" />)
+    fireEvent.error(container.querySelector("img") as HTMLImageElement)
+    expect(container.querySelector("img")).toBeNull()
+
+    rerender(<SearchSiteLogo site="instagram" />)
+
+    expect(container.querySelector("img")).not.toBeNull()
+  })
+})
diff --git a/surfsense_local/frontend/src/features/sources/web/search-site-logo.tsx b/surfsense_local/frontend/src/features/sources/web/search-site-logo.tsx
index 63fd0f4089..5037e1e9a3 100644
--- a/surfsense_local/frontend/src/features/sources/web/search-site-logo.tsx
+++ b/surfsense_local/frontend/src/features/sources/web/search-site-logo.tsx
@@ -20,6 +20,8 @@ const SITE_PLUGIN = {
 export function SearchSiteLogo({ site }: { site: SearchSite }) {
   return (
     <PluginIcon
+      // A logo that failed says nothing about the next site's.
+      key={site}
       id={SITE_PLUGIN[site]}
       name={searchSiteName(site)}
       hasIcon