@cartanova/qgrid-cli 2.10.10 → 2.11.1

package/bundle/src/application/qgrid/qgrid.frame.test.ts

@@ -1,11 +1,39 @@
 import { EventEmitter } from "node:events";
 
-import { Sonamu } from "sonamu";
+import Fastify from "fastify";
+import { BadRequestException, ServiceUnavailableException, Sonamu } from "sonamu";
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 
-import { QgridFrame } from "./qgrid.frame";
-import { QueryInput } from "./qgrid.types";
+import { SD } from "../../i18n/sd.generated";
+import { handleServerError } from "../../server-error-handler";
+import { mapRequestBodyTooLarge, QgridFrame } from "./qgrid.frame";
+import { QgridDispatcher, QgridDispatcherClass } from "./qgrid.dispatcher";
+import {
+  GenerateImagesInput,
+  MEDIA_MAX_INLINE_BYTES,
+  QueryInput,
+  REQUEST_BODY_LIMIT_BYTES,
+  TranscribeInput,
+  VideoStartInput,
+} from "./qgrid.types";
+import {
+  GROK_OAUTH_CLIENT_ID,
+  GROK_OAUTH_SCOPES,
+  type GrokCredentials,
+  GrokCapacityError,
+  GrokInvalidRequestError,
+  GrokNoWorkersError,
+  GrokOAuthError,
+  GrokOnDemandExcludedError,
+  GrokRateLimitedError,
+  GrokZdrVideoUnavailableError,
+} from "../../utils/providers/grok";
 import { ImageGenerationError } from "../../utils/providers/openai/openai-dispatcher";
+import type {
+  GenerateRequest,
+  GenerateResult,
+  GenerateStreamCallbacks,
+} from "../../utils/providers/common/provider-dispatcher";
 
 const {
   findOneMock,
@@ -18,6 +46,7 @@ const {
   requestLogCreateRunMock,
   requestLogFinishRunMock,
   appendStepMock,
+  continueToolRunMock,
   dispatcherQueryMock,
   dispatcherQueryStreamMock,
   beforeQueryMock,
@@ -25,6 +54,10 @@ const {
   finishRunWithErrorMock,
   finishRunAbortedMock,
   getRateLimitsByTokenIdMock,
+  replaceByAccountMock,
+  completeGrokOAuthMock,
+  startGrokDeviceCodeMock,
+  pollGrokDeviceCodeMock,
 } = vi.hoisted(() => {
   // isolate:false 공유 레지스트리 방어: 같은 fork 에서 앞선 파일이 request-log/token/
   // run-lifecycle 실모듈을 캐시해 두면 아래 vi.mock 들이 조용히 무시된다(실 DB 호출로
@@ -41,6 +74,7 @@ const {
     requestLogCreateRunMock: vi.fn(),
     requestLogFinishRunMock: vi.fn(),
     appendStepMock: vi.fn(),
+    continueToolRunMock: vi.fn(),
     dispatcherQueryMock: vi.fn(),
     dispatcherQueryStreamMock: vi.fn(),
     beforeQueryMock: vi.fn(),
@@ -48,9 +82,37 @@ const {
     finishRunWithErrorMock: vi.fn(),
     finishRunAbortedMock: vi.fn(),
     getRateLimitsByTokenIdMock: vi.fn(),
+    replaceByAccountMock: vi.fn(),
+    completeGrokOAuthMock: vi.fn(),
+    startGrokDeviceCodeMock: vi.fn(),
+    pollGrokDeviceCodeMock: vi.fn(),
   };
 });
 
+// OAuth pending state lives in the cache; a Map stub supports the round trip.
+vi.mock("sonamu/cache", () => {
+  const store = new Map<string, string>();
+  return {
+    getCacheManagerRef: () => ({
+      set: async ({ key, value }: { key: string; value: string }) => {
+        store.set(key, value);
+      },
+      get: async ({ key }: { key: string }) => store.get(key),
+      delete: async ({ key }: { key: string }) => {
+        store.delete(key);
+      },
+    }),
+  };
+});
+
+// Token exchange and device endpoints hit xAI; startGrokOAuth stays real (fetch is stubbed).
+vi.mock("../../utils/providers/grok", async (importOriginal) => ({
+  ...(await importOriginal<typeof import("../../utils/providers/grok")>()),
+  completeGrokOAuth: completeGrokOAuthMock,
+  startGrokDeviceCode: startGrokDeviceCodeMock,
+  pollGrokDeviceCode: pollGrokDeviceCodeMock,
+}));
+
 vi.mock("../request-log/request-log.model", () => ({
   MICRO_USD: 1_000_000,
   RequestLogModel: {
@@ -58,6 +120,7 @@ vi.mock("../request-log/request-log.model", () => ({
     createRun: requestLogCreateRunMock,
     finishRun: requestLogFinishRunMock,
     appendStep: appendStepMock,
+    continueToolRun: continueToolRunMock,
   },
 }));
 
@@ -69,6 +132,7 @@ vi.mock("../token/token.model", () => ({
     toggleActive: toggleActiveMock,
     save: saveMock,
     updateFields: updateFieldsMock,
+    replaceByAccount: replaceByAccountMock,
   },
 }));
 
@@ -89,6 +153,10 @@ vi.mock("./qgrid.dispatcher", async (importOriginal) => {
       openaiDispatcher: {
         getRateLimitsByTokenId: getRateLimitsByTokenIdMock,
       },
+      grokDispatcher: null,
+      startupState: { openai: "starting", anthropic: "starting" },
+      tokens: new Map(),
+      subscriber: null,
     },
   };
 });
@@ -1427,3 +1495,945 @@ describe("QgridFrame.usage", () => {
     });
   });
 });
+
+function grokJwtSegment(value: unknown): string {
+  return Buffer.from(JSON.stringify(value)).toString("base64url");
+}
+
+function grokCredentials(overrides: Partial<GrokCredentials> = {}): GrokCredentials {
+  return {
+    authMode: "oauth",
+    accessToken: `${grokJwtSegment({ alg: "ES256", typ: "at+jwt" })}.${grokJwtSegment({
+      iss: "https://auth.x.ai",
+      sub: "11111111-0000-4000-8000-000000000001",
+      exp: 2_000_000_000,
+    })}.c2ln`,
+    refreshToken: "grok-refresh-1",
+    expiresAt: "2030-01-01T00:00:00Z",
+    issuer: "https://auth.x.ai",
+    clientId: GROK_OAUTH_CLIENT_ID,
+    scopes: [...GROK_OAUTH_SCOPES],
+    userId: "11111111-0000-4000-8000-000000000001",
+    sessionId: "22222222-0000-4000-8000-000000000001",
+    agentId: "33333333-0000-4000-8000-000000000001",
+    subscriptionTier: "Super",
+    ...overrides,
+  };
+}
+
+describe("QgridFrame Grok registration and routing", () => {
+  beforeEach(() => {
+    // Discovery fails offline, so startGrokOAuth uses the fixed xAI endpoints.
+    vi.stubGlobal(
+      "fetch",
+      vi.fn(async () => {
+        throw new Error("offline");
+      }),
+    );
+    replaceByAccountMock.mockReset().mockResolvedValue([1]);
+    completeGrokOAuthMock.mockReset();
+    startGrokDeviceCodeMock.mockReset();
+    pollGrokDeviceCodeMock.mockReset();
+    findManyMock.mockReset();
+    beforeQueryMock.mockReset().mockResolvedValue({ requestLogId: 51, stepIndex: 0 });
+    afterQueryMock.mockReset().mockResolvedValue({});
+    dispatcherQueryMock.mockReset();
+  });
+
+  afterEach(() => {
+    vi.unstubAllGlobals();
+    Object.assign(QgridDispatcher, { grokDispatcher: null });
+    delete QgridDispatcher.startupState.grok;
+  });
+
+  async function startGrokLogin(name: string) {
+    const result = await QgridFrame.oauthStartGrok(name);
+    const url = new URL(result.authUrl);
+    return { result, url, state: url.searchParams.get("state") ?? "" };
+  }
+
+  it("oauthStartGrok returns an xAI authorize URL with the Grok client id in paste mode", async () => {
+    const { result, url, state } = await startGrokLogin("grok-a");
+
+    expect(result.mode).toBe("code");
+    expect(url.origin).toBe("https://auth.x.ai");
+    expect(url.searchParams.get("client_id")).toBe(GROK_OAUTH_CLIENT_ID);
+    expect(url.searchParams.get("code_challenge_method")).toBe("S256");
+    expect(url.searchParams.get("redirect_uri")).toMatch(/^http:\/\/127\.0\.0\.1:\d+\/callback$/);
+    expect(state.length).toBeGreaterThan(20);
+  });
+
+  it("oauthComplete binds a bare code to the pending Grok login and stores Grok credentials", async () => {
+    const { url, state } = await startGrokLogin("grok-a");
+    const credentials = grokCredentials();
+    completeGrokOAuthMock.mockResolvedValueOnce(credentials);
+
+    await expect(QgridFrame.oauthComplete("  bare-code-123  ")).resolves.toEqual({
+      added: true,
+      name: "grok-a",
+    });
+
+    expect(completeGrokOAuthMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        codeOrUrl: "  bare-code-123  ",
+        state,
+        redirectUri: url.searchParams.get("redirect_uri"),
+        verifier: expect.any(String),
+      }),
+    );
+    expect(replaceByAccountMock).toHaveBeenCalledWith("grok", credentials.userId, {
+      provider: "grok",
+      credentials,
+      name: "grok-a",
+    });
+    // The pending login is single-use.
+    await expect(QgridFrame.oauthComplete("bare-code-123")).rejects.toThrow("code#state");
+  });
+
+  it("oauthComplete accepts the pasted callback URL by its state", async () => {
+    const { url, state } = await startGrokLogin("grok-b");
+    completeGrokOAuthMock.mockResolvedValueOnce(grokCredentials());
+    const pasted = `${url.searchParams.get("redirect_uri")}?code=abc&state=${state}`;
+
+    await expect(QgridFrame.oauthComplete(pasted)).resolves.toEqual({ added: true, name: "grok-b" });
+    expect(completeGrokOAuthMock).toHaveBeenCalledWith(
+      expect.objectContaining({ codeOrUrl: pasted, state }),
+    );
+  });
+
+  it("oauthComplete rejects an exchange that does not yield an at+jwt", async () => {
+    await startGrokLogin("grok-a");
+    completeGrokOAuthMock.mockRejectedValueOnce(
+      new GrokOAuthError("invalid_access_token", "access token is not an at+jwt"),
+    );
+
+    await expect(QgridFrame.oauthComplete("bare-code")).rejects.toThrow(
+      SD("qgrid.grokLoginFailed")("access token is not an at+jwt"),
+    );
+    expect(replaceByAccountMock).not.toHaveBeenCalled();
+  });
+
+  it("refuses to store API-key shaped credentials", async () => {
+    await startGrokLogin("grok-a");
+    completeGrokOAuthMock.mockResolvedValueOnce(grokCredentials({ accessToken: "xai-api-key" }));
+
+    await expect(QgridFrame.oauthComplete("bare-code")).rejects.toThrow(
+      SD("qgrid.grokNotOAuthToken"),
+    );
+    expect(replaceByAccountMock).not.toHaveBeenCalled();
+  });
+
+  it("rejects a bare code when no Grok login is pending", async () => {
+    await expect(QgridFrame.oauthComplete("orphan-code")).rejects.toThrow("code#state");
+    expect(completeGrokOAuthMock).not.toHaveBeenCalled();
+  });
+
+  it("device code login polls pending, slow_down, then added", async () => {
+    startGrokDeviceCodeMock.mockResolvedValueOnce({
+      deviceCode: "dev-1",
+      userCode: "ABCD-EFGH",
+      verificationUri: "https://accounts.x.ai/device",
+      expiresIn: 600,
+      interval: 5,
+    });
+    const credentials = grokCredentials();
+    pollGrokDeviceCodeMock
+      .mockResolvedValueOnce({ status: "pending", interval: 5 })
+      .mockResolvedValueOnce({ status: "pending", interval: 10 })
+      .mockResolvedValueOnce({ status: "complete", credentials });
+
+    await expect(QgridFrame.oauthStartGrokDevice("grok-dev")).resolves.toMatchObject({
+      userCode: "ABCD-EFGH",
+      interval: 5,
+    });
+    await expect(QgridFrame.oauthPollGrokDevice("dev-1")).resolves.toEqual({
+      status: "pending",
+      intervalSec: 5,
+    });
+    await expect(QgridFrame.oauthPollGrokDevice("dev-1")).resolves.toEqual({
+      status: "slow_down",
+      intervalSec: 10,
+    });
+    await expect(QgridFrame.oauthPollGrokDevice("dev-1")).resolves.toEqual({
+      status: "added",
+      name: "grok-dev",
+    });
+
+    expect(pollGrokDeviceCodeMock).toHaveBeenLastCalledWith({ deviceCode: "dev-1", interval: 10 });
+    expect(replaceByAccountMock).toHaveBeenCalledWith("grok", credentials.userId, {
+      provider: "grok",
+      credentials,
+      name: "grok-dev",
+    });
+    await expect(QgridFrame.oauthPollGrokDevice("dev-1")).rejects.toThrow("expired");
+  });
+
+  it("device code polling drops the login when the user denies it", async () => {
+    startGrokDeviceCodeMock.mockResolvedValueOnce({
+      deviceCode: "dev-2",
+      userCode: "WXYZ",
+      verificationUri: "https://accounts.x.ai/device",
+      expiresIn: 600,
+      interval: 5,
+    });
+    pollGrokDeviceCodeMock.mockRejectedValueOnce(
+      new GrokOAuthError("access_denied", "authorization denied by the user"),
+    );
+
+    await QgridFrame.oauthStartGrokDevice("grok-dev");
+    await expect(QgridFrame.oauthPollGrokDevice("dev-2")).rejects.toThrow("denied");
+    await expect(QgridFrame.oauthPollGrokDevice("dev-2")).rejects.toThrow("expired");
+    expect(pollGrokDeviceCodeMock).toHaveBeenCalledTimes(1);
+  });
+
+  it("query with grok/<model> reaches the registered Grok dispatcher", async () => {
+    const generate = vi.fn(async () => ({
+      text: "grok says hi",
+      tokenName: "grok-a",
+      usage: {
+        totalTokens: 3,
+        inputTokens: 2,
+        cachedInputTokens: 0,
+        outputTokens: 1,
+        reasoningOutputTokens: 0,
+      },
+      durationMs: 5,
+      model: "grok-4.7",
+      threadCoord: { workerId: 1, threadId: "t", epoch: 0 },
+    }));
+    const routing = new QgridDispatcherClass({
+      grokDispatcher: {
+        generate,
+        generateStream: vi.fn(),
+        start: vi.fn(async () => {}),
+        stop: vi.fn(async () => {}),
+        onTokenAdded: vi.fn(),
+        onTokenUpdated: vi.fn(),
+        onTokenRemoved: vi.fn(),
+        replaceTokens: vi.fn(),
+      },
+    });
+    dispatcherQueryMock.mockImplementationOnce(
+      (input: Parameters<typeof routing.query>[0], signal?: AbortSignal) =>
+        routing.query(input, signal),
+    );
+
+    const result = await QgridFrame.query({ prompt: "hi", model: "grok/grok-4.7" });
+
+    expect(generate).toHaveBeenCalledWith(expect.objectContaining({ model: "grok-4.7" }));
+    expect(result.text).toBe("grok says hi");
+  });
+
+  it("usage reports cached Grok billing without keepalive", async () => {
+    const credentials = grokCredentials();
+    findManyMock.mockResolvedValueOnce({
+      rows: [{ ...tokenEntry, id: 9, provider: "grok", name: "grok-a", credentials }],
+    });
+    Object.assign(QgridDispatcher, {
+      grokDispatcher: {
+        tokenStates: () => [
+          {
+            id: 9,
+            name: "grok-a",
+            lastBilling: {
+              utilizationPct: 42,
+              periodType: "USAGE_PERIOD_TYPE_WEEKLY",
+              periodEnd: "2030-01-08T00:00:00Z",
+              fetchedAt: 1,
+            },
+            onDemandExcluded: true,
+            zdrVideoExcluded: true,
+            lastTier: "Ultra",
+          },
+        ],
+      },
+    });
+
+    await expect(QgridFrame.usage(9)).resolves.toEqual({
+      provider: "grok",
+      fiveHour: null,
+      sevenDay: { utilization: 42, resetsAt: "2030-01-08T00:00:00Z" },
+      periodType: "USAGE_PERIOD_TYPE_WEEKLY",
+      tier: "Ultra",
+      onDemandExcluded: true,
+      zdrVideoExcluded: true,
+      keepaliveApplicable: false,
+    });
+  });
+
+  it("usage falls back to the stored tier before the first billing fetch", async () => {
+    findManyMock.mockResolvedValueOnce({
+      rows: [
+        { ...tokenEntry, id: 9, provider: "grok", name: "grok-a", credentials: grokCredentials() },
+      ],
+    });
+
+    await expect(QgridFrame.usage(9)).resolves.toMatchObject({
+      provider: "grok",
+      sevenDay: null,
+      tier: "Super",
+      onDemandExcluded: false,
+      zdrVideoExcluded: false,
+      keepaliveApplicable: false,
+    });
+  });
+
+  it("health includes the Grok startup state", async () => {
+    QgridDispatcher.startupState.grok = "ready";
+
+    const health = await QgridFrame.health();
+
+    expect(health.providers.grok).toBe("ready");
+    expect(health.ready).toBe(true);
+  });
+});
+
+
+// ─── U11: Grok media/voice endpoints, providerOptions/providerMetadata, tool emulation ───
+
+type FakeGrok = Record<string, ReturnType<typeof vi.fn>>;
+
+function grokTurn(overrides: Partial<GenerateResult> = {}): GenerateResult {
+  return {
+    text: "grok says hi",
+    tokenName: "grok/a",
+    usage: {
+      totalTokens: 30,
+      inputTokens: 20,
+      cachedInputTokens: 5,
+      outputTokens: 10,
+      reasoningOutputTokens: 2,
+      contextTokens: 30,
+    },
+    durationMs: 5,
+    ttftMs: 2,
+    model: "grok-4.7",
+    threadCoord: { workerId: 7, threadId: "t", epoch: -1 },
+    costUsd: 0.000_338,
+    costSource: "provider" as const,
+    providerMetadata: { xai: { costInUsdTicks: 3_380_000, contextWindow: 256_000 } },
+    ...overrides,
+  };
+}
+
+function directProvider(result: GenerateResult) {
+  return {
+    generate: vi.fn(async (_req: GenerateRequest) => result),
+    generateStream: vi.fn(async (_req: GenerateRequest, cb: GenerateStreamCallbacks) => {
+      cb.onDelta(result.text);
+      cb.onComplete(result);
+    }),
+    start: vi.fn(async () => {}),
+    stop: vi.fn(async () => {}),
+    onTokenAdded: vi.fn(),
+    onTokenUpdated: vi.fn(),
+    onTokenRemoved: vi.fn(),
+    replaceTokens: vi.fn(),
+  };
+}
+
+const PNG_B64 = Buffer.from([0x89, 0x50, 0x4e, 0x47, 1, 2, 3, 4]).toString("base64");
+
+describe("QgridFrame Grok media and voice (U11)", () => {
+  let grok: FakeGrok;
+
+  beforeEach(() => {
+    requestLogCreateRunMock.mockReset().mockResolvedValue(77);
+    requestLogFinishRunMock.mockReset().mockResolvedValue(undefined);
+    appendStepMock.mockReset().mockResolvedValue(1);
+    continueToolRunMock.mockReset().mockResolvedValue(1);
+    findActiveByProviderAndNameMock.mockReset();
+    grok = {
+      generateImages: vi.fn(),
+      editImages: vi.fn(),
+      startVideo: vi.fn(),
+      getVideo: vi.fn(),
+      transcribe: vi.fn(),
+      speak: vi.fn(),
+    };
+    Object.assign(QgridDispatcher, { grokDispatcher: grok });
+  });
+
+  afterEach(() => {
+    Object.assign(QgridDispatcher, { grokDispatcher: null });
+  });
+
+  it("generateImages routes grok/* to the Grok dispatcher and returns b64 plus the reported cost", async () => {
+    grok.editImages!.mockResolvedValueOnce({
+      result: {
+        data: [{ b64_json: PNG_B64, mime_type: "image/png", revised_prompt: "a cat" }],
+        usage: { cost_in_usd_ticks: 500_000_000 },
+      },
+      tokenId: 3,
+      tokenName: "grok/a",
+    });
+    findActiveByProviderAndNameMock.mockResolvedValueOnce({ id: 3 });
+    const ref = "data:image/png;base64," + PNG_B64;
+
+    const out = await QgridFrame.generateImages({
+      model: "grok/grok-imagine-image-quality",
+      prompt: "cat",
+      n: 1,
+      aspectRatio: "1:1",
+      referenceImages: [ref],
+      tokenName: "grok/a",
+      projectName: "proj",
+    });
+
+    expect(grok.generateImages).not.toHaveBeenCalled();
+    expect(grok.editImages).toHaveBeenCalledWith(
+      expect.objectContaining({
+        model: "grok-imagine-image-quality",
+        prompt: "cat",
+        n: 1,
+        aspectRatio: "1:1",
+        images: [ref],
+        preferredTokenId: 3,
+        requirePreferredToken: true,
+      }),
+    );
+    expect(out).toEqual({
+      model: "grok/grok-imagine-image-quality",
+      tokenName: "grok/a",
+      images: [{ b64: PNG_B64, mimeType: "image/png", revisedPrompt: "a cat" }],
+      usage: { imageCount: 1, costInUsdTicks: 500_000_000 },
+      costUsd: 0.05,
+      costSource: "provider",
+      requestLogId: 77,
+    });
+    expect(requestLogCreateRunMock).toHaveBeenCalledWith(
+      expect.objectContaining({
+        requested_model_name: "grok/grok-imagine-image-quality",
+        project_name: "proj",
+        is_image_generation: true,
+      }),
+    );
+    const step = appendStepMock.mock.calls[0]![1] as Record<string, unknown>;
+    expect(step).toMatchObject({ type: "tool_call", tool_name: "image_generation" });
+    expect(JSON.parse(step.tool_args as string)).toMatchObject({
+      kind: "image",
+      imageCount: 1,
+      costInUsdTicks: 500_000_000,
+      costSource: "provider",
+      inputImages: [{ mediaType: "image/png", data: PNG_B64 }],
+    });
+    expect(requestLogFinishRunMock).toHaveBeenCalledWith(
+      77,
+      expect.objectContaining({
+        status: "succeeded",
+        token_name: "grok/a",
+        model_name: "grok/grok-imagine-image-quality",
+        cost_usd: 0,
+        cost_source: "provider",
+        is_image_generation: true,
+        image_cost_usd: 50_000,
+        image_cost_method: "reported:grok/grok-imagine-image-quality:ticks=500000000",
+      }),
+    );
+  });
+
+  it("generateImages falls back to the catalog estimate when xAI reports no ticks", async () => {
+    grok.generateImages!.mockResolvedValueOnce({
+      result: { data: [{ b64_json: PNG_B64 }, { b64_json: PNG_B64 }] },
+      tokenId: 3,
+      tokenName: "grok/a",
+    });
+
+    const out = await QgridFrame.generateImages({ model: "grok/grok-imagine-image-quality", prompt: "x", n: 2, logger: false });
+
+    expect(out.costUsd).toBeCloseTo(0.1, 12);
+    expect(out.costSource).toBe("pricing_table");
+    expect(out.images.map((image) => image.mimeType)).toEqual(["image/png", "image/png"]);
+    expect(out.requestLogId).toBeUndefined();
+    expect(requestLogCreateRunMock).not.toHaveBeenCalled();
+  });
+
+  it.each([
+    [{ model: "openai/gpt-image-2", prompt: "x" }, "model"],
+    [{ model: "grok/grok-imagine-image", prompt: "x", referenceImages: ["http://example.com/a.png"] }, "https"],
+    [{ model: "grok/grok-imagine-image", prompt: "x", referenceImages: ["data:image/gif;base64,R0lG"] }, "image/png"],
+    [
+      {
+        model: "grok/grok-imagine-image",
+        prompt: "x",
+        referenceImages: ["data:image/png;base64," + "A".repeat(28 * 1024 * 1024)],
+      },
+      "7 MiB",
+    ],
+  ])("generateImages rejects invalid input %#", async (input, message) => {
+    await expect(QgridFrame.generateImages(input)).rejects.toThrow(message);
+    expect(grok.generateImages).not.toHaveBeenCalled();
+    expect(grok.editImages).not.toHaveBeenCalled();
+  });
+
+  it("caps inline media so every accepted request fits the JSON body limit", () => {
+    // Largest base64 string whose decoded size is exactly the inline budget.
+    const b64 = (bytes: number) => Buffer.alloc(bytes).toString("base64");
+    const atLimit = "data:image/png;base64," + b64(MEDIA_MAX_INLINE_BYTES);
+    const overLimit = "data:image/png;base64," + b64(MEDIA_MAX_INLINE_BYTES + 1);
+    const half = "data:image/png;base64," + b64(MEDIA_MAX_INLINE_BYTES / 2 + 1);
+
+    const largest = { model: "grok/grok-imagine-image", prompt: "x".repeat(4096), referenceImages: [atLimit] };
+    expect(GenerateImagesInput.safeParse(largest).success).toBe(true);
+    expect(Buffer.byteLength(JSON.stringify(largest))).toBeLessThan(REQUEST_BODY_LIMIT_BYTES);
+    expect(GenerateImagesInput.safeParse({ ...largest, referenceImages: [overLimit] }).success).toBe(false);
+    // The aggregate counts every inline image of the request, not just each one.
+    const split = GenerateImagesInput.safeParse({ ...largest, referenceImages: [half, half] });
+    expect(split.success).toBe(false);
+    expect(split.error?.issues.map((issue) => issue.message)).toContain("inline images exceed 7 MiB in total");
+    expect(
+      VideoStartInput.safeParse({ model: "grok/grok-imagine-video", prompt: "x", image: half, lastFrame: half }).success,
+    ).toBe(false);
+
+    // qgrid has no source-video input, and reference-to-video needs reference images.
+    const video = { model: "grok/grok-imagine-video", prompt: "x" };
+    for (const mode of ["edit-video", "extend-video"]) {
+      expect(VideoStartInput.safeParse({ ...video, mode }).success).toBe(false);
+    }
+    const noReferences = VideoStartInput.safeParse({ ...video, mode: "reference-to-video" });
+    expect(noReferences.error?.issues.map((issue) => issue.path.join("."))).toEqual(["referenceImages"]);
+    expect(
+      VideoStartInput.safeParse({ ...video, mode: "reference-to-video", referenceImages: ["https://x/r.png"] })
+        .success,
+    ).toBe(true);
+
+    const audio = { audio: b64(MEDIA_MAX_INLINE_BYTES), mediaType: "audio/wav" };
+    expect(TranscribeInput.safeParse(audio).success).toBe(true);
+    expect(Buffer.byteLength(JSON.stringify(audio))).toBeLessThan(REQUEST_BODY_LIMIT_BYTES);
+    expect(TranscribeInput.safeParse({ ...audio, audio: b64(MEDIA_MAX_INLINE_BYTES + 1) }).success).toBe(false);
+  });
+
+  it("answers an oversize JSON body with a Sonamu 413 instead of a 500", async () => {
+    const server = Fastify({ bodyLimit: REQUEST_BODY_LIMIT_BYTES });
+    server.setErrorHandler((error, _request, reply) => {
+      handleServerError(mapRequestBodyTooLarge(error), reply);
+    });
+    server.post("/media", async () => ({ ok: true }));
+    try {
+      const payload = JSON.stringify({ audio: "A".repeat(REQUEST_BODY_LIMIT_BYTES), mediaType: "audio/wav" });
+      const tooLarge = await server.inject({
+        method: "POST",
+        url: "/media",
+        headers: { "content-type": "application/json" },
+        payload,
+      });
+      expect(tooLarge.statusCode).toBe(413);
+      expect(tooLarge.json()).toEqual({ name: "Error", message: SD("qgrid.requestBodyTooLarge")(10) });
+
+      const ok = await server.inject({ method: "POST", url: "/media", payload: { ok: true } });
+      expect(ok.statusCode).toBe(200);
+    } finally {
+      await server.close();
+    }
+  });
+
+  it("videoStart answers GrokZdrVideoUnavailableError as 503 and logs the provider message", async () => {
+    const zdr = new GrokZdrVideoUnavailableError(["grok/a"]);
+    grok.startVideo!.mockRejectedValueOnce(zdr);
+
+    const error = await QgridFrame.videoStart({ model: "grok/grok-imagine-video-1.5", prompt: "x" }).catch((e: unknown) => e);
+
+    expect(error).toBeInstanceOf(ServiceUnavailableException);
+    expect(error).toMatchObject({ statusCode: 503, message: SD("qgrid.grokZdrVideoUnavailable")("grok/a") });
+    expect(requestLogFinishRunMock).toHaveBeenCalledWith(77, { status: "error", error_message: zdr.message });
+  });
+
+  it("generateImages records a failed provider call as an error run", async () => {
+    grok.generateImages!.mockRejectedValueOnce(new Error("upstream 500"));
+
+    await expect(QgridFrame.generateImages({ model: "grok/grok-imagine-image", prompt: "x" })).rejects.toThrow("upstream 500");
+
+    expect(requestLogFinishRunMock).toHaveBeenCalledWith(77, { status: "error", error_message: "upstream 500" });
+  });
+
+  it("videoStart/videoStatus polls the starting token, checks the owner, and passes the URL through", async () => {
+    grok.startVideo!.mockResolvedValueOnce({ result: { request_id: "req-1" }, tokenId: 9, tokenName: "grok/b" });
+    const url = "https://vidgen.x.ai/abc/video.mp4?sig=opaque";
+    grok.getVideo!
+      .mockResolvedValueOnce({ result: { status: "pending" }, tokenId: 9, tokenName: "grok/b" })
+      .mockResolvedValueOnce({
+        result: { status: "done", video: { url, duration: 6 }, usage: { cost_in_usd_ticks: 2_000_000_000 } },
+        tokenId: 9,
+        tokenName: "grok/b",
+      });
+
+    const started = await QgridFrame.videoStart({
+      model: "grok/grok-imagine-video-1.5",
+      prompt: "waves",
+      duration: 6,
+      resolution: "720p",
+      projectName: "proj",
+    });
+    expect(started.jobHandle).toMatch(/^[A-Za-z0-9_-]{22}$/);
+    expect(started).toMatchObject({ model: "grok/grok-imagine-video-1.5", requestLogId: 77 });
+    expect(grok.startVideo).toHaveBeenCalledWith(
+      expect.objectContaining({ model: "grok-imagine-video-1.5", duration: 6, resolution: "720p" }),
+    );
+    expect(appendStepMock).toHaveBeenCalledWith(
+      77,
+      expect.objectContaining({ type: "tool_call", tool_name: "video_generation", tool_call_id: "video_generation:req-1" }),
+    );
+    expect(appendStepMock.mock.calls[0]![1]).not.toHaveProperty("tool_result");
+
+    // A different owner, or no owner, sees the same 404 as an unknown handle.
+    await expect(QgridFrame.videoStatus({ jobHandle: started.jobHandle, projectName: "other" })).rejects.toThrow(SD("qgrid.videoJobNotFound"));
+    await expect(QgridFrame.videoStatus({ jobHandle: started.jobHandle })).rejects.toThrow(SD("qgrid.videoJobNotFound"));
+    await expect(QgridFrame.videoStatus({ jobHandle: "A".repeat(22), projectName: "proj" })).rejects.toThrow(SD("qgrid.videoJobNotFound"));
+    expect(grok.getVideo).not.toHaveBeenCalled();
+
+    await expect(QgridFrame.videoStatus({ jobHandle: started.jobHandle, projectName: "proj" })).resolves.toEqual({
+      status: "pending",
+      mimeType: "video/mp4",
+      requestLogId: 77,
+    });
+    const done = await QgridFrame.videoStatus({ jobHandle: started.jobHandle, projectName: "proj" });
+    expect(done).toEqual({
+      status: "done",
+      url,
+      mimeType: "video/mp4",
+      durationSec: 6,
+      costInUsdTicks: 2_000_000_000,
+      costUsd: 0.2,
+      requestLogId: 77,
+    });
+    expect(grok.getVideo).toHaveBeenCalledWith(9, "req-1", expect.any(AbortSignal));
+
+    // The terminal result is cached: no further upstream poll, no second log write.
+    await expect(QgridFrame.videoStatus({ jobHandle: started.jobHandle, projectName: "proj" })).resolves.toEqual(done);
+    expect(grok.getVideo).toHaveBeenCalledTimes(2);
+    expect(continueToolRunMock).toHaveBeenCalledTimes(1);
+    expect(continueToolRunMock).toHaveBeenCalledWith(77, [
+      expect.objectContaining({ toolCallId: "video_generation:req-1", isError: false }),
+    ]);
+    expect(requestLogFinishRunMock).toHaveBeenCalledTimes(1);
+    expect(requestLogFinishRunMock).toHaveBeenCalledWith(
+      77,
+      expect.objectContaining({
+        status: "succeeded",
+        response: url,
+        token_name: "grok/b",
+        model_name: "grok/grok-imagine-video-1.5",
+        cost_usd: 200_000,
+        cost_source: "provider",
+      }),
+    );
+  });
+
+  it("videoStatus logs a terminal result once when polls overlap", async () => {
+    grok.startVideo!.mockResolvedValueOnce({ result: { request_id: "req-1" }, tokenId: 9, tokenName: "grok/b" });
+    let release!: () => void;
+    const upstream = new Promise<void>((resolve) => (release = resolve));
+    grok.getVideo!.mockImplementation(async () => {
+      await upstream;
+      return { result: { status: "done", video: { url: "https://x/v.mp4" } }, tokenId: 9, tokenName: "grok/b" };
+    });
+    const { jobHandle } = await QgridFrame.videoStart({ model: "grok/grok-imagine-video-1.5", prompt: "x" });
+
+    const polls = [QgridFrame.videoStatus({ jobHandle }), QgridFrame.videoStatus({ jobHandle })];
+    release();
+    const [first, second] = await Promise.all(polls);
+
+    expect(grok.getVideo).toHaveBeenCalledTimes(2);
+    expect(second).toEqual(first);
+    expect(continueToolRunMock).toHaveBeenCalledTimes(1);
+    expect(requestLogFinishRunMock).toHaveBeenCalledTimes(1);
+  });
+
+  it("videoStatus reports a content-policy block as a failed job", async () => {
+    grok.startVideo!.mockResolvedValueOnce({ result: { request_id: "req-2" }, tokenId: 9, tokenName: "grok/b" });
+    grok.getVideo!.mockResolvedValueOnce({
+      result: { status: "done", video: { url: "https://x/v.mp4", respect_moderation: false } },
+      tokenId: 9,
+      tokenName: "grok/b",
+    });
+    const { jobHandle } = await QgridFrame.videoStart({ model: "grok/grok-imagine-video-1.5", prompt: "x" });
+
+    const status = await QgridFrame.videoStatus({ jobHandle });
+
+    expect(status).toMatchObject({ status: "failed", mimeType: "video/mp4" });
+    expect(status.url).toBeUndefined();
+    expect(requestLogFinishRunMock).toHaveBeenCalledWith(77, expect.objectContaining({ status: "error" }));
+  });
+
+  it("transcribe decodes base64 audio and returns the transcript", async () => {
+    grok.transcribe!.mockResolvedValueOnce({
+      result: { text: "hello there", language: "en", duration: 1.5 },
+      tokenId: 3,
+      tokenName: "grok/a",
+    });
+    const audio = Buffer.from("RIFFfakewav").toString("base64");
+
+    const out = await QgridFrame.transcribe({ audio, mediaType: "audio/wav", language: "en" });
+
+    expect(out).toEqual({
+      text: "hello there",
+      language: "en",
+      durationSec: 1.5,
+      model: "grok/grok-voice-transcribe-2.0",
+      tokenName: "grok/a",
+      requestLogId: 77,
+    });
+    const request = grok.transcribe!.mock.calls[0]![0] as { audio: Uint8Array; mediaType: string; model: string };
+    expect(Buffer.from(request.audio).toString()).toBe("RIFFfakewav");
+    expect(request).toMatchObject({ mediaType: "audio/wav", model: "grok-voice-transcribe-2.0" });
+    expect(requestLogFinishRunMock).toHaveBeenCalledWith(
+      77,
+      expect.objectContaining({ status: "succeeded", response: "hello there", cost_usd: null, cost_source: null }),
+    );
+  });
+
+  it("transcribe rejects non-base64 and non-audio input before dispatch", async () => {
+    await expect(QgridFrame.transcribe({ audio: "not base64!", mediaType: "audio/wav" })).rejects.toThrow("base64");
+    await expect(QgridFrame.transcribe({ audio: "AAAA", mediaType: "video/mp4" })).rejects.toThrow("audio/*");
+    expect(grok.transcribe).not.toHaveBeenCalled();
+  });
+
+  it("speak applies the /tts defaults and returns base64 audio", async () => {
+    const bytes = new Uint8Array([0xff, 0xfb, 1, 2]);
+    grok.speak!.mockResolvedValueOnce({ result: { audio: bytes, mediaType: "audio/mpeg" }, tokenId: 3, tokenName: "grok/a" });
+
+    const out = await QgridFrame.speak({ text: "hi" });
+
+    expect(grok.speak).toHaveBeenCalledWith(
+      expect.objectContaining({ text: "hi", voice: "ara", language: "en", format: "mp3" }),
+    );
+    expect(out).toEqual({
+      audio: Buffer.from(bytes).toString("base64"),
+      mimeType: "audio/mpeg",
+      tokenName: "grok/a",
+      requestLogId: 77,
+    });
+    expect(Buffer.from(out.audio, "base64")).toEqual(Buffer.from(bytes));
+    // No ticks reported: the cost is unknown, not a guessed 0.
+    expect(requestLogFinishRunMock).toHaveBeenCalledWith(
+      77,
+      expect.objectContaining({ status: "succeeded", cost_usd: null, cost_source: null }),
+    );
+  });
+
+  it("media endpoints answer not-ready when no Grok media dispatcher is registered", async () => {
+    Object.assign(QgridDispatcher, { grokDispatcher: null });
+    await expect(QgridFrame.speak({ text: "hi" })).rejects.toThrow(SD("qgrid.dispatcherStarting")("Grok"));
+    expect(requestLogCreateRunMock).not.toHaveBeenCalled();
+  });
+});
+
+describe("QgridFrame Grok query contract (U11)", () => {
+  beforeEach(() => {
+    beforeQueryMock.mockReset().mockResolvedValue({ requestLogId: 61, stepIndex: 0 });
+    afterQueryMock.mockReset().mockResolvedValue({});
+    dispatcherQueryMock.mockReset();
+    dispatcherQueryStreamMock.mockReset();
+  });
+
+  afterEach(() => {
+    vi.restoreAllMocks();
+  });
+
+  it("forwards providerOptions.xai and surfaces providerMetadata.xai with the provider cost", async () => {
+    const grok = directProvider(grokTurn());
+    const routing = new QgridDispatcherClass({ grokDispatcher: grok });
+    dispatcherQueryMock.mockImplementationOnce((input: Parameters<typeof routing.query>[0], signal?: AbortSignal) =>
+      routing.query(input, signal),
+    );
+
+    const out = await QgridFrame.query({
+      prompt: "hi",
+      model: "grok/grok-4.7",
+      providerOptions: { xai: { serviceTier: "priority", parallelToolCalls: false } },
+    });
+
+    expect(grok.generate).toHaveBeenCalledWith(
+      expect.objectContaining({
+        model: "grok-4.7",
+        providerOptions: { xai: { serviceTier: "priority", parallelToolCalls: false } },
+      }),
+    );
+    expect(out).toMatchObject({
+      requestLogId: 61,
+      costUsd: 0.000_338,
+      costSource: "provider",
+      providerMetadata: { xai: { costInUsdTicks: 3_380_000, contextWindow: 256_000 } },
+    });
+  });
+
+  it("keeps a reported zero-tick cost authoritative and prices a tickless turn from the catalog", async () => {
+    const zero = directProvider(grokTurn({ costUsd: 0, providerMetadata: { xai: { costInUsdTicks: 0 } } }));
+    const zeroOut = await new QgridDispatcherClass({ grokDispatcher: zero }).query({ prompt: "hi", model: "grok/grok-4.7" });
+    expect(zeroOut).toMatchObject({ costUsd: 0, costSource: "provider" });
+
+    const tickless = directProvider(grokTurn({ costUsd: undefined, costSource: undefined, providerMetadata: undefined }));
+    const ticklessOut = await new QgridDispatcherClass({ grokDispatcher: tickless }).query({ prompt: "hi", model: "grok/grok-4.7" });
+    // 15 uncached input * $2/M + 5 cached * $0.5/M + 10 output * $6/M
+    expect(ticklessOut.costSource).toBe("pricing_table");
+    expect(ticklessOut.costUsd).toBeCloseTo((15 * 2 + 5 * 0.5 + 10 * 6) / 1_000_000, 15);
+    expect(ticklessOut.providerMetadata).toBeUndefined();
+  });
+
+  it("publishes providerMetadata.xai on the stream done event", async () => {
+    const grok = directProvider(grokTurn());
+    const routing = new QgridDispatcherClass({ grokDispatcher: grok });
+    dispatcherQueryStreamMock.mockImplementationOnce(
+      (input: Parameters<typeof routing.queryStream>[0], cb: Parameters<typeof routing.queryStream>[1], signal?: AbortSignal) =>
+        routing.queryStream(input, cb, signal),
+    );
+    const sse = { closed: false, publish: vi.fn(), onClose: vi.fn(), end: vi.fn(async () => {}) };
+    vi.spyOn(Sonamu, "getContext").mockReturnValue({ createSSE: vi.fn(() => sse) } as never);
+
+    const { streamId } = await QgridFrame.prepareStream({
+      prompt: "hi",
+      model: "grok/grok-4.7",
+      providerOptions: { xai: { serviceTier: "default" } },
+    });
+    await QgridFrame.queryStream(streamId);
+
+    expect(grok.generateStream).toHaveBeenCalledWith(
+      expect.objectContaining({ providerOptions: { xai: { serviceTier: "default" } } }),
+      expect.anything(),
+    );
+    expect(sse.publish).toHaveBeenCalledWith(
+      "done",
+      expect.objectContaining({
+        requestLogId: 61,
+        costSource: "provider",
+        providerMetadata: { xai: { costInUsdTicks: 3_380_000, contextWindow: 256_000 } },
+      }),
+    );
+  });
+
+  it.each([true, false])(
+    "answers a Grok caller-validation error as 400, logged run=%s",
+    async (logged) => {
+      const grok = directProvider(grokTurn());
+      const invalid = new GrokInvalidRequestError("Unsupported Grok history item at index 0");
+      grok.generate.mockRejectedValueOnce(invalid);
+      const routing = new QgridDispatcherClass({ grokDispatcher: grok });
+      dispatcherQueryMock.mockImplementationOnce((input: Parameters<typeof routing.query>[0], signal?: AbortSignal) =>
+        routing.query(input, signal),
+      );
+
+      const error = await QgridFrame.query({
+        prompt: "hi",
+        model: "grok/grok-4.7",
+        history: "[{}]",
+        ...(logged ? {} : { logger: false }),
+      }).catch((e: unknown) => e);
+
+      expect(error).toBeInstanceOf(BadRequestException);
+      expect(error).toMatchObject({
+        statusCode: 400,
+        message: SD("qgrid.grokInvalidRequest")(invalid.message),
+      });
+      if (logged) {
+        // The run keeps the provider's own message.
+        expect(finishRunWithErrorMock).toHaveBeenCalledWith(61, invalid.message, expect.anything());
+      }
+    },
+  );
+
+  it("keeps Grok upstream failures as server errors", async () => {
+    const grok = directProvider(grokTurn());
+    const upstream = new Error("grok upstream 502");
+    grok.generate.mockRejectedValueOnce(upstream);
+    const routing = new QgridDispatcherClass({ grokDispatcher: grok });
+    dispatcherQueryMock.mockImplementationOnce((input: Parameters<typeof routing.query>[0], signal?: AbortSignal) =>
+      routing.query(input, signal),
+    );
+
+    await expect(QgridFrame.query({ prompt: "hi", model: "grok/grok-4.7" })).rejects.toBe(upstream);
+  });
+
+  it.each([
+    [new GrokOnDemandExcludedError(["grok-a", "grok-b"]), SD("qgrid.grokOnDemandExcluded")("grok-a, grok-b")],
+    [new GrokCapacityError(), SD("qgrid.grokCapacity")],
+    [new GrokNoWorkersError(), SD("qgrid.grokNoWorkers")("NO_GROK_WORKERS")],
+  ])("answers a Grok no-eligible-token error as 503 (%s)", async (excluded, message) => {
+    const grok = directProvider(grokTurn());
+    grok.generate.mockRejectedValueOnce(excluded);
+    const routing = new QgridDispatcherClass({ grokDispatcher: grok });
+    dispatcherQueryMock.mockImplementationOnce((input: Parameters<typeof routing.query>[0], signal?: AbortSignal) =>
+      routing.query(input, signal),
+    );
+
+    const error = await QgridFrame.query({ prompt: "hi", model: "grok/grok-4.7" }).catch((e: unknown) => e);
+
+    expect(error).toBeInstanceOf(ServiceUnavailableException);
+    expect(error).toMatchObject({ statusCode: 503, message });
+  });
+
+  it("answers GrokRateLimitedError as 429 with the cooldown in Retry-After", async () => {
+    const header = vi.fn();
+    const responseRaw = Object.assign(new EventEmitter(), { headersSent: false, writableEnded: false });
+    vi.spyOn(Sonamu, "getContext").mockReturnValue({
+      transport: "http",
+      request: { raw: Object.assign(new EventEmitter(), { aborted: false }) },
+      reply: { raw: responseRaw, header },
+    } as never);
+    const grok = directProvider(grokTurn());
+    grok.generate.mockRejectedValueOnce(new GrokRateLimitedError(42));
+    const routing = new QgridDispatcherClass({ grokDispatcher: grok });
+    dispatcherQueryMock.mockImplementationOnce((input: Parameters<typeof routing.query>[0], signal?: AbortSignal) =>
+      routing.query(input, signal),
+    );
+
+    const error = await QgridFrame.query({ prompt: "hi", model: "grok/grok-4.7", logger: false }).catch(
+      (e: unknown) => e,
+    );
+
+    expect(error).toMatchObject({ statusCode: 429, message: SD("qgrid.grokRateLimited")(42) });
+    expect(header).toHaveBeenCalledWith("Retry-After", "42");
+
+    // The server error handler keeps a Sonamu 429 and does not pin a 503 Retry-After over it.
+    const reply = { header: vi.fn(), status: vi.fn(), send: vi.fn() };
+    reply.status.mockReturnValue(reply);
+    handleServerError(error as Error, reply as never);
+    expect(reply.status).toHaveBeenCalledWith(429);
+    expect(reply.header).not.toHaveBeenCalled();
+  });
+
+  it("runs grok tools through the same structured-output emulation as openai (D8)", async () => {
+    const tools = [
+      {
+        name: "get_weather",
+        description: "weather",
+        inputSchema: { type: "object", properties: { city: { type: "string" } }, required: ["city"] },
+      },
+    ];
+    const envelope = JSON.stringify({
+      result: { action: "tool_call", answer: null, toolCalls: [{ toolName: "get_weather", args: '{"city":"Seoul"}' }] },
+    });
+    const grok = directProvider(grokTurn({ text: envelope }));
+    const openai = directProvider(grokTurn({ text: envelope, model: "gpt-6-sol", providerMetadata: undefined }));
+    const routing = new QgridDispatcherClass({ grokDispatcher: grok });
+    routing.openaiDispatcher = openai as never;
+
+    const grokOut = await routing.query({ prompt: "weather?", model: "grok/grok-4.7", tools });
+    const openaiOut = await routing.query({ prompt: "weather?", model: "openai/gpt-6-sol", tools });
+
+    const grokReq = grok.generate.mock.calls[0]![0];
+    const openaiReq = openai.generate.mock.calls[0]![0];
+    // No native tool plumbing: the envelope schema is the only tool contract on the wire.
+    expect(grokReq.tools).toBeUndefined();
+    expect(grokReq.outputSchema).toEqual(openaiReq.outputSchema);
+    expect(typeof grokReq.stopAfterOutputMessage).toBe("function");
+    for (const out of [grokOut, openaiOut]) {
+      expect(out.finishReason).toBe("tool-calls");
+      expect(out.content).toEqual([
+        { type: "tool-call", toolCallId: expect.stringMatching(/^call_/), toolName: "get_weather", input: '{"city":"Seoul"}' },
+      ]);
+    }
+
+    // Follow-up with tool results continues the existing way and decodes the answer envelope.
+    const answer = JSON.stringify({ result: { action: "answer", answer: "Sunny", toolCalls: null } });
+    grok.generate.mockResolvedValueOnce(grokTurn({ text: answer }));
+    const followUp = await routing.query({
+      prompt: "weather?",
+      model: "grok/grok-4.7",
+      tools,
+      toolResults: [{ toolCallId: "call_1", toolName: "get_weather", output: "sunny, 21C" }],
+    });
+    expect(followUp).toMatchObject({ finishReason: "stop", text: "Sunny" });
+    const followReq = grok.generate.mock.calls[1]![0];
+    expect(JSON.stringify(followReq.coldInput)).toContain("sunny, 21C");
+  });
+});