improve(ui): verify llama.cpp model setup

docs/plugins/llama-cpp.md

@@ -25,22 +25,23 @@ deleting a manually installed runtime inside the OpenClaw package directory.
 
 ## Local text inference
 
-Choose **Local model (llama.cpp)** during interactive onboarding. OpenClaw asks
-before downloading the default model:
+Choose **llama.cpp** during interactive onboarding. OpenClaw installs the
+official provider plugin, then asks before downloading the default model:
 
-`hf:bartowski/Qwen_Qwen3-4B-Instruct-2507-GGUF/Qwen_Qwen3-4B-Instruct-2507-Q4_K_M.gguf`
+`hf:unsloth/gemma-4-E4B-it-GGUF/gemma-4-E4B-it-Q4_K_M.gguf`
 
-The Qwen3 4B Instruct 2507 Q4_K_M file is about 2.5 GB. Budget roughly 3 GB of
-RAM for model weights, plus context and OpenClaw runtime overhead. The default
-context is automatically sized with an 8,192-token cap so it remains practical
-on 8 GB machines. Configure a larger context only when the machine has enough
+The Gemma 4 E4B IT Q4_K_M file is about 5.0 GB. OpenClaw offers this default
+only on machines with at least 16 GiB of RAM, leaving room for model weights,
+context, and Gateway overhead. The default context is automatically sized with
+an 8,192-token cap. Configure a larger context only when the machine has enough
 memory.
 
 The onboarding discovery check is read-only. It offers llama.cpp automatically
 only when the default or configured GGUF file is already in the model cache; it
 never downloads during discovery. Ollama and LM Studio remain separate local
 service choices and keep their own discovery flows. Manually choosing llama.cpp
-is the path that prompts for the default model download.
+is the path that installs the runtime, prompts for the default model download,
+and verifies a real model reply before marking setup complete.
 
 The provider uses the GGUF model's embedded chat template and native
 node-llama-cpp function calling. Text streams token by token. Tool calls return

extensions/llama-cpp/README.md

@@ -16,7 +16,7 @@ native installs and updates.
 
 ## Configure text inference
 
-Choose **Local model (llama.cpp)** during onboarding. After explicit consent,
+Choose **llama.cpp** during onboarding. After explicit consent,
 OpenClaw downloads Gemma 4 E4B IT Q4_K_M (approximately 5.0 GB) as the default.
 The bundled download is offered only on machines with at least 16 GiB of RAM.
 Discovery never downloads a model.

ui/public/provider-icons/ATTRIBUTION.md

@@ -17,6 +17,39 @@ content otherwise match the cited upstream files.
 
 Provider names and marks remain the property of their respective owners.
 
+## llama.cpp icon
+
+`ProviderIcon-llamacpp.svg` is a metadata-cleaned copy of
+`media/llama1-icon-transparent.svg` from the official llama.cpp repository at
+commit `2cfef4d117d67ab1dec002915b48a15d11ee1973`:
+
+- Source:
+  https://github.com/ggml-org/llama.cpp/blob/2cfef4d117d67ab1dec002915b48a15d11ee1973/media/llama1-icon-transparent.svg
+- Upstream license:
+  https://github.com/ggml-org/llama.cpp/blob/master/LICENSE
+
+MIT License
+
+Copyright (c) 2023-2026 The ggml authors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
 ## CodexBar license
 
 MIT License

ui/public/provider-icons/ProviderIcon-llamacpp.svg

@@ -0,0 +1,8 @@
+<svg viewBox="0 0 250 250" xmlns="http://www.w3.org/2000/svg">
+  <g transform="translate(-995.51066 -129.70875)" fill="white">
+    <path d="m1163.3 226.8-13.5 24c-17.8-13.7-44.2-15.7-62-1-28.7 23.7-26.7 78.5 18 78.8 12.5 0 23.1-5.9 34.5-9.8l6 23.9c-10.1 4.7-20.4 9.5-31.5 11-101.2 13.8-95.4-132.3-3.9-139.9 19.2-1.6 36.1 3.4 52.5 13z"/>
+    <path d="M1093.4 203.8c-15.4 4.6-29.7 13.1-40.5 25-2-24.2 3.4-73.1 30.3-82.7 4-1.4 17.7-4.9 17.3 2.2-.4 7.1-9.9 19.3-12.2 25.9-4 11.6-.3 19.6 5.2 29.7z"/>
+    <path d="M1131.4 307.8h-15v-17h-17v-14h15.5l1.5-1.5v-16.5h15v18h16v14h-16zm55 0h-15v-17h-16v-14h15.5l1.5-1.5v-16.5h15v16.5l1.5 1.5h15.5v14h-17z"/>
+    <path d="M1142.3 156.9c2 3-9.3 15.9-11.1 19.2-5.2 9.8-1.7 15.4 2.2 24.7-11.3-1.7-21.8-.3-33 1 2.5-21.5 14.6-52.8 41.9-44.9z"/>
+  </g>
+</svg>

ui/src/components/provider-icon.ts

@@ -37,6 +37,7 @@ const PROVIDER_ICON_NAMES = new Set([
   "kimi",
   "kiro",
   "litellm",
+  "llamacpp",
   "llmproxy",
   "manus",
   "mimo",
@@ -71,6 +72,7 @@ const PROVIDER_ICON_ALIASES: Readonly<Record<string, string>> = {
   google: "gemini",
   "google-gemini-cli": "gemini",
   "github-copilot": "copilot",
+  "llama-cpp": "llamacpp",
   // CodexBar names its bundled OpenAI knot asset "codex".
   openai: "codex",
   moonshot: "kimi",
@@ -87,6 +89,7 @@ const PROVIDER_DISPLAY_LABELS: Readonly<Record<string, string>> = {
   anthropic: "Anthropic",
   google: "Google",
   "github-copilot": "GitHub",
+  "llama-cpp": "llama.cpp",
   openai: "OpenAI",
   moonshot: "Moonshot AI",
   opencode: "OpenCode",

ui/src/components/wizard-step-controls.ts

@@ -80,6 +80,15 @@ function renderContinueStep(props: WizardStepControlsProps) {
   `;
 }
 
+function renderProgressStep(step: WizardStep) {
+  return html`
+    <div class="wizard-step__progress" role="status" aria-live="polite">
+      <span class="wizard-step__spinner" aria-hidden="true"></span>
+      ${renderMessage(step)}
+    </div>
+  `;
+}
+
 function renderTextStep(props: WizardStepControlsProps) {
   const step = props.step;
   const value = typeof props.value === "string" ? props.value : "";
@@ -221,10 +230,12 @@ export function renderWizardStepControls(
       return renderConfirmStep(props);
     case "multiselect":
       return renderMultiselectStep(props);
-    // These carry no input of their own: they show whatever the step supplies
-    // (message, link, device code) behind a single Continue.
-    case "note":
     case "progress":
+      return props.step.executor === "gateway"
+        ? renderProgressStep(props.step)
+        : renderContinueStep(props);
+    // These show whatever the step supplies behind a single Continue.
+    case "note":
     case "action":
       return renderContinueStep(props);
   }

ui/src/e2e/model-setup.e2e.test.ts

@@ -345,10 +345,29 @@ describeControlUiE2e("Control UI Model Setup mocked Gateway E2E", () => {
 
       const start = await gateway.waitForRequest("openclaw.setup.prepare.start");
       expect(start.params).toMatchObject({ authChoice: "ollama" });
+
+      if (artifactDir) {
+        await mkdir(artifactDir, { recursive: true });
+        await page.screenshot({
+          animations: "disabled",
+          fullPage: true,
+          path: path.join(artifactDir, "ollama-mode-select-desktop.png"),
+        });
+      }
+
       await page.getByRole("radio", { name: /Local only/u }).check();
       await page.getByRole("button", { name: "Continue" }).click();
       const baseUrl = page.getByLabel("Ollama base URL");
       await expect.poll(() => baseUrl.inputValue()).toBe("http://127.0.0.1:11434");
+
+      if (artifactDir) {
+        await page.screenshot({
+          animations: "disabled",
+          fullPage: true,
+          path: path.join(artifactDir, "ollama-host-desktop.png"),
+        });
+      }
+
       await page.getByRole("button", { name: "Submit" }).click();
       await page.getByText("Ollama could not be reached at http://127.0.0.1:11434.").waitFor();
       await page
@@ -413,6 +432,167 @@ describeControlUiE2e("Control UI Model Setup mocked Gateway E2E", () => {
     }
   });
 
+  it("downloads, verifies, and opens chat with llama.cpp", async () => {
+    const context = await browser.newContext({
+      colorScheme: "dark",
+      locale: "en-US",
+      serviceWorkers: "block",
+      viewport: { height: 900, width: 1280 },
+    });
+    const page = await context.newPage();
+    const initialDetection = {
+      candidates: [],
+      manualProviders: [],
+      workspace: "/tmp/openclaw-e2e",
+      setupComplete: false,
+    };
+    const modelRef = "llama-cpp/gemma-4-e4b-it-q4_k_m";
+    const gateway = await installMockGateway(page, {
+      featureMethods: [
+        "chat.metadata",
+        "chat.startup",
+        "openclaw.setup.detect",
+        "openclaw.setup.activate",
+        "openclaw.setup.prepare.start",
+        "wizard.next",
+      ],
+      methodResponses: {
+        "openclaw.setup.detect": initialDetection,
+        "openclaw.setup.prepare.start": {
+          sessionId: "llama-cpp-prepare-session",
+          done: false,
+          status: "running",
+        },
+        "openclaw.setup.activate": {
+          ok: true,
+          modelRef,
+          latencyMs: 731,
+          lines: ["Model ready"],
+        },
+        "wizard.next": {
+          sequence: [
+            {
+              done: false,
+              status: "running",
+              step: {
+                id: "llama-cpp-consent",
+                type: "confirm",
+                message:
+                  "Download Gemma 4 E4B IT Q4_K_M (about 5.0 GB) for local llama.cpp inference?",
+                initialValue: false,
+              },
+            },
+            {
+              done: false,
+              status: "running",
+              step: {
+                id: "llama-cpp-download-20",
+                type: "progress",
+                message: "Downloading Gemma 4 E4B… 20% (1.0/5.0 GB, 38 MB/s)",
+                executor: "gateway",
+              },
+            },
+            {
+              done: false,
+              status: "running",
+              step: {
+                id: "llama-cpp-download-100",
+                type: "progress",
+                message: "Gemma 4 E4B model downloaded",
+                executor: "gateway",
+              },
+            },
+            { done: true, status: "done" },
+          ],
+        },
+      },
+    });
+
+    try {
+      const response = await page.goto(`${server.baseUrl}settings/model-setup`);
+      expect(response?.status()).toBe(200);
+      const llamaCppRow = page.locator('[data-prepare-choice="llama-cpp"]');
+      await llamaCppRow.getByRole("button", { name: "Check & set up" }).waitFor();
+      await expect
+        .poll(() => llamaCppRow.locator('[data-provider-icon="llamacpp"]').count())
+        .toBe(1);
+      await expect.poll(() => llamaCppRow.textContent()).not.toContain("Gemma");
+      await expect.poll(() => llamaCppRow.textContent()).not.toContain("GB");
+      await expect.poll(() => llamaCppRow.textContent()).not.toContain("RAM");
+
+      if (artifactDir) {
+        await mkdir(artifactDir, { recursive: true });
+        await page.screenshot({
+          animations: "disabled",
+          fullPage: true,
+          path: path.join(artifactDir, "llama-cpp-offer-desktop.png"),
+        });
+      }
+
+      await llamaCppRow.getByRole("button", { name: "Check & set up" }).click();
+      const start = await gateway.waitForRequest("openclaw.setup.prepare.start");
+      expect(start.params).toMatchObject({ authChoice: "llama-cpp" });
+      await page.getByRole("heading", { name: "Set up a local model" }).waitFor();
+      await page.getByText("Download Gemma 4 E4B IT Q4_K_M").waitFor();
+
+      if (artifactDir) {
+        await page.screenshot({
+          animations: "disabled",
+          fullPage: true,
+          path: path.join(artifactDir, "llama-cpp-confirm-desktop.png"),
+        });
+      }
+
+      await gateway.setMethodResponse("openclaw.setup.detect", {
+        ...initialDetection,
+        candidates: [
+          {
+            kind: "provider-auto:llama-cpp",
+            brandId: "llama-cpp",
+            label: "llama.cpp",
+            detail: "Gemma 4 E4B downloaded",
+            modelRef,
+            recommended: true,
+            credentials: true,
+          },
+        ],
+      });
+      await page.getByRole("button", { name: "Yes" }).click();
+      await page.getByRole("heading", { name: "Connection verified" }).waitFor();
+      await expect
+        .poll(() => page.locator(".model-setup-success").textContent())
+        .toContain(modelRef);
+      await expect
+        .poll(() => page.locator(".model-setup-success").textContent())
+        .toContain("Verified in 731 ms");
+
+      const activate = await gateway.waitForRequest("openclaw.setup.activate");
+      expect(activate.params).toEqual({
+        kind: "provider-auto:llama-cpp",
+        modelRef,
+      });
+
+      if (artifactDir) {
+        await page.screenshot({
+          animations: "disabled",
+          fullPage: true,
+          path: path.join(artifactDir, "llama-cpp-ready-desktop.png"),
+        });
+        await page.setViewportSize({ height: 844, width: 390 });
+        await page.screenshot({
+          animations: "disabled",
+          fullPage: true,
+          path: path.join(artifactDir, "llama-cpp-ready-mobile.png"),
+        });
+      }
+
+      await page.getByRole("button", { name: "Start chatting" }).click();
+      await expect.poll(() => new URL(page.url()).pathname).toBe("/chat");
+    } finally {
+      await context.close();
+    }
+  });
+
   it("turns an unverifiable Gemini CLI login into direct recovery actions", async () => {
     const context = await browser.newContext({
       colorScheme: "dark",

ui/src/i18n/locales/en.ts

@@ -2021,15 +2021,14 @@ export const en: TranslationMap = {
       more: "More sign-in options",
     },
     prepare: {
-      title: "Set up a local model",
-      intro:
-        "OpenClaw checks the local service, confirms tool support, and helps prepare a compatible model.",
-      button: "Set up / Download model",
+      title: "Run a model locally",
+      intro: "Use a local model service, or run a private GGUF model directly inside this Gateway.",
       ollamaButton: "Check & set up",
       ollamaLabel: "Ollama",
       ollamaHint: "Connect to the Ollama service on this Gateway and prepare a tools-capable model",
-      llamaCppLabel: "Local model (llama.cpp)",
-      llamaCppHint: "Download an approximately 5.0 GB local model; requires 16 GB RAM",
+      llamaCppLabel: "llama.cpp",
+      llamaCppNotReady:
+        "llama.cpp did not produce a usable local model. Review the setup result, then retry.",
     },
     manual: {
       title: "Connect with an API key or token",

ui/src/pages/model-setup/model-setup-page.test.ts

@@ -277,6 +277,114 @@ describe("ModelSetupPage catalog icons", () => {
     });
   });
 
+  it("verifies a prepared llama.cpp model before showing success", async () => {
+    const { context: baseContext, client, request } = createContext();
+    const runtimeConfig = {
+      runExternalMutation: vi.fn(async (task) => ({
+        ok: true as const,
+        value: await task(client),
+        refresh: { ok: true as const },
+      })),
+    } as unknown as ApplicationContext["runtimeConfig"];
+    const context = { ...baseContext, runtimeConfig } as ApplicationContext;
+    request.mockImplementation(async (method: string) => {
+      if (method === "openclaw.setup.prepare.start") {
+        return { sessionId: "prepare-session", done: false, status: "running" };
+      }
+      if (method === "wizard.next") {
+        return { done: true, status: "done" };
+      }
+      if (method === "openclaw.setup.detect") {
+        return {
+          ...detection,
+          candidates: [
+            {
+              kind: "existing-model",
+              label: "Existing llama.cpp model",
+              detail: "Already configured",
+              modelRef: "llama-cpp/custom",
+              recommended: false,
+              credentials: true,
+            },
+            {
+              kind: "provider-auto:llama-cpp",
+              brandId: "llama-cpp",
+              label: "llama.cpp",
+              detail: "Gemma 4 E4B downloaded",
+              modelRef: "llama-cpp/gemma-4-e4b-it-q4_k_m",
+              recommended: true,
+              credentials: true,
+            },
+          ],
+        };
+      }
+      if (method === "openclaw.setup.activate") {
+        return {
+          ok: true,
+          modelRef: "llama-cpp/gemma-4-e4b-it-q4_k_m",
+          latencyMs: 731,
+          lines: ["Model ready"],
+        };
+      }
+      return {};
+    });
+    const { page } = await mountPage(context, {
+      state: { phase: "ready", result: detection },
+      client,
+      firstRun: false,
+    });
+
+    page.querySelector<HTMLButtonElement>('[data-prepare-choice="llama-cpp"] button')?.click();
+
+    await vi.waitFor(() => {
+      expect(request).toHaveBeenCalledWith(
+        "openclaw.setup.activate",
+        {
+          kind: "provider-auto:llama-cpp",
+          modelRef: "llama-cpp/gemma-4-e4b-it-q4_k_m",
+        },
+        expect.objectContaining({ signal: expect.any(AbortSignal) }),
+      );
+      expect(page.textContent).toContain("Connection verified");
+      expect(page.textContent).toContain("llama-cpp/gemma-4-e4b-it-q4_k_m");
+      expect(page.textContent).toContain("Verified in 731 ms");
+    });
+  });
+
+  it("keeps an incomplete llama.cpp setup visible instead of claiming success", async () => {
+    const { context, client, request } = createContext();
+    request.mockImplementation(async (method: string) => {
+      if (method === "openclaw.setup.prepare.start") {
+        return { sessionId: "prepare-session", done: false, status: "running" };
+      }
+      if (method === "wizard.next") {
+        return { done: true, status: "done" };
+      }
+      if (method === "openclaw.setup.detect") {
+        return detection;
+      }
+      return {};
+    });
+    const { page } = await mountPage(context, {
+      state: { phase: "ready", result: detection },
+      client,
+      firstRun: false,
+    });
+
+    page.querySelector<HTMLButtonElement>('[data-prepare-choice="llama-cpp"] button')?.click();
+
+    await vi.waitFor(() => {
+      expect(page.textContent).toContain(
+        "llama.cpp did not produce a usable local model. Review the setup result, then retry.",
+      );
+    });
+    expect(request).not.toHaveBeenCalledWith(
+      "openclaw.setup.activate",
+      expect.anything(),
+      expect.anything(),
+    );
+  });
+
   it("flushes a pending config draft before one-shot activation and refreshes afterward", async () => {
     vi.useFakeTimers();
     const { context, client, request, runtimeConfig } = createContext();

ui/src/pages/model-setup/model-setup-page.ts

@@ -18,7 +18,7 @@ import { isGatewayMethodAdvertised } from "../../lib/gateway-methods.ts";
 import { OpenClawLightDomElement } from "../../lit/openclaw-element.ts";
 import { SubscriptionsController } from "../../lit/subscriptions-controller.ts";
 import { fetchCatalogIconBlobUrl } from "../plugins/icon-loader.ts";
-import type { ModelSetupPrepareOption } from "./prepare-options.ts";
+import { findPreparedModelCandidate, type ModelSetupPrepareOption } from "./prepare-options.ts";
 import { detectModelSetup, verifyModelSetup } from "./rpc.ts";
 import {
   activationTargetId,
@@ -93,6 +93,7 @@ export class ModelSetupPage extends OpenClawLightDomElement {
 
   private observedClient: GatewayBrowserClient | null = null;
   private dataClient: GatewayBrowserClient | null = null;
+  private pendingPrepareOption: ModelSetupPrepareOption | null = null;
   private readonly iconMisses = new Set<string>();
   private readonly iconRequests = new Map<
     string,
@@ -511,6 +512,9 @@ export class ModelSetupPage extends OpenClawLightDomElement {
   }
 
   private async handleWizardDone(startMethod: ModelSetupWizardStartMethod): Promise<void> {
+    const prepareOption =
+      startMethod === "openclaw.setup.prepare.start" ? this.pendingPrepareOption : null;
+    this.pendingPrepareOption = null;
     const result = await this.detect();
     if (!result) {
       this.wizard.fail(t("modelSetup.errors.requestFailed"));
@@ -526,6 +530,26 @@ export class ModelSetupPage extends OpenClawLightDomElement {
         modelRef: result.configuredModel ?? t("modelSetup.success.configuredModel"),
       };
     }
+    if (prepareOption?.activateAfterPrepare) {
+      const candidate = findPreparedModelCandidate(result, prepareOption.id);
+      if (!candidate) {
+        this.wizard.fail(t("modelSetup.prepare.llamaCppNotReady"));
+        return;
+      }
+      this.wizard.close();
+      this.activateCandidate(candidate);
+      return;
+    }
+    this.wizard.close();
+  }
+
+  private cancelWizard(): void {
+    this.pendingPrepareOption = null;
+    void this.wizard.cancel();
+  }
+
+  private closeWizard(): void {
+    this.pendingPrepareOption = null;
     this.wizard.close();
   }
 
@@ -574,10 +598,12 @@ export class ModelSetupPage extends OpenClawLightDomElement {
       onVerify: () => void this.verifyConnection(),
       onActivateCandidate: (candidate) => this.activateCandidate(candidate),
       onStartAuth: (option: AuthOption) => {
+        this.pendingPrepareOption = null;
         this.wizardMode = "auth";
         void this.wizard.start(option.id);
       },
       onStartPrepare: (option: ModelSetupPrepareOption) => {
+        this.pendingPrepareOption = option;
         this.wizardMode = "prepare";
         void this.wizard.start(option.id, "openclaw.setup.prepare.start");
       },
@@ -603,8 +629,8 @@ export class ModelSetupPage extends OpenClawLightDomElement {
       },
       onWizardValueChange: (value) => (this.wizardValue = value),
       onWizardAnswer: (value, includeValue) => void this.wizard.answer(value, includeValue),
-      onWizardCancel: () => void this.wizard.cancel(),
-      onWizardClose: () => this.wizard.close(),
+      onWizardCancel: () => this.cancelWizard(),
+      onWizardClose: () => this.closeWizard(),
     });
     return html`
       <section class="content-header">

ui/src/pages/model-setup/prepare-options.ts

@@ -6,6 +6,7 @@ export type ModelSetupPrepareOption = {
   brandId?: string;
   label: string;
   hint?: string;
+  activateAfterPrepare?: boolean;
   icon?: string;
   website?: string;
 };
@@ -24,7 +25,9 @@ export function listModelSetupPrepareOptions(
       id: "llama-cpp",
       brandId: "llama-cpp",
       label: t("modelSetup.prepare.llamaCppLabel"),
-      hint: t("modelSetup.prepare.llamaCppHint"),
+      // Keep model choice and resource requirements in the provider-owned
+      // consent step so local runtimes remain peers in this list.
+      activateAfterPrepare: true,
     },
   ];
   const presented = [
@@ -46,3 +49,10 @@ export function listModelSetupPrepareOptions(
       return wire ? Object.assign({}, choice, wire, { id: choice.id }) : choice;
     });
 }
+
+export function findPreparedModelCandidate(
+  result: SystemAgentSetupDetectResult,
+  providerId: ModelSetupPrepareOption["id"],
+) {
+  return result.candidates.find((candidate) => candidate.kind === `provider-auto:${providerId}`);
+}

ui/src/pages/model-setup/view.test.ts

@@ -182,7 +182,7 @@ describe("renderModelSetup", () => {
     expect(text(container)).toContain("Found, but needs attention");
     expect(text(container)).toContain("OpenClaw could not confirm a usable login");
     expect(text(container)).toContain("Sign in with a provider");
-    expect(text(container)).toContain("Set up a local model");
+    expect(text(container)).toContain("Run a model locally");
     expect(text(container)).toContain("Connect with an API key or token");
     expect(
       container.querySelector('[data-manual-provider="openai"][data-selected]'),
@@ -496,7 +496,12 @@ describe("renderModelSetup", () => {
       '[data-prepare-choice="llama-cpp"] button',
     );
     expect(ollama?.textContent).toContain("Check & set up");
-    expect(llamaCpp).not.toBeNull();
+    expect(llamaCpp?.textContent).toContain("Check & set up");
+    const llamaCppRow = container.querySelector('[data-prepare-choice="llama-cpp"]');
+    expect(llamaCppRow?.querySelector('[data-provider-icon="llamacpp"]')).not.toBeNull();
+    expect(text(llamaCppRow!)).toContain("llama.cpp");
+    expect(text(llamaCppRow!)).not.toContain("Gemma");
+    expect(llamaCppRow?.classList.contains("model-setup__prepare-row--featured")).toBe(false);
     ollama?.click();
     expect(onStartPrepare).toHaveBeenCalledWith(expect.objectContaining({ id: "ollama" }));
 
@@ -991,7 +996,7 @@ describe("renderModelSetup", () => {
     expect(text(confirm)).toContain("No");
   });
 
-  it.each(["multiselect", "progress", "action"] as const)("renders the %s wizard step", (type) => {
+  it.each(["multiselect", "action"] as const)("renders the %s wizard step", (type) => {
     const container = wizardStep({
       id: type,
       type,
@@ -1003,4 +1008,31 @@ describe("renderModelSetup", () => {
     expect(text(container)).toContain(`${type} message`);
     expect(text(container)).toContain("Continue");
   });
+
+  it("renders gateway progress without an answer control", () => {
+    const container = wizardStep({
+      id: "download",
+      type: "progress",
+      message: "Downloading Gemma 4 E4B… 42%",
+      executor: "gateway",
+    });
+
+    expect(text(container)).toContain("Downloading Gemma 4 E4B… 42%");
+    expect(container.querySelector('[role="status"]')).not.toBeNull();
+    expect(container.querySelector(".wizard-step__spinner")).not.toBeNull();
+    expect(container.querySelector(".wizard-step__progress button")).toBeNull();
+  });
+
+  it("keeps a Continue action for client progress", () => {
+    const container = wizardStep({
+      id: "client-progress",
+      type: "progress",
+      message: "Waiting for the local client",
+      executor: "client",
+    });
+
+    expect(text(container)).toContain("Waiting for the local client");
+    expect(text(container)).toContain("Continue");
+    expect(container.querySelector('[role="status"]')).toBeNull();
+  });
 });

ui/src/pages/model-setup/view.ts

@@ -429,9 +429,7 @@ function renderPrepare(props: ModelSetupViewProps, result: SystemAgentSetupDetec
                 ?disabled=${props.actionsDisabled}
                 @click=${() => props.onStartPrepare(option)}
               >
-                ${option.id === "ollama"
-                  ? t("modelSetup.prepare.ollamaButton")
-                  : t("modelSetup.prepare.button")}
+                ${t("modelSetup.prepare.ollamaButton")}
               </button>
             </div>
           `,

ui/src/styles/components.css

@@ -5747,6 +5747,10 @@ td.data-table-key-col {
   color: #4285f4;
 }
 
+.provider-brand-icon[data-provider-icon="llamacpp"] {
+  color: #ff8236;
+}
+
 .provider-brand-icon[data-provider-icon="opencode"],
 .provider-brand-icon[data-provider-icon="opencodego"] {
   color: var(--text-strong);

ui/src/styles/wizard-step-controls.css

@@ -48,3 +48,37 @@
   display: flex;
   gap: 8px;
 }
+
+.wizard-step__progress {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  min-height: 44px;
+  padding: 12px 14px;
+  border: 1px solid var(--border);
+  border-radius: var(--radius-md);
+  background: var(--bg);
+}
+
+.wizard-step__spinner {
+  width: 18px;
+  height: 18px;
+  flex: 0 0 18px;
+  border: 2px solid color-mix(in srgb, var(--accent) 24%, transparent);
+  border-top-color: var(--accent);
+  border-radius: 50%;
+  animation: wizard-step-spin 800ms linear infinite;
+}
+
+@keyframes wizard-step-spin {
+  to {
+    transform: rotate(360deg);
+  }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  .wizard-step__spinner {
+    animation: none;
+    border-color: var(--accent);
+  }
+}