fix(web): stop dashboard grounded-copy AI truncating at its output cap

Co-authored-by: polylane[bot] <277585245+polylane[bot]@users.noreply.github.com>

Author: stewartjarod

apps/web/src/lib/ai/__tests__/grounded-copy-request.test.ts

@@ -0,0 +1,71 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { z } from "zod";
+
+const { mockGetAIModel, mockGenerateText } = vi.hoisted(() => ({
+  mockGetAIModel: vi.fn(),
+  mockGenerateText: vi.fn(),
+}));
+
+vi.mock("@wraps/ai", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("@wraps/ai")>();
+  // Partial mock: keep DEPLOYMENT_MODE_ENV and the rest that `aiEnv()` reads.
+  return { ...actual, getAIModel: mockGetAIModel };
+});
+
+vi.mock("ai", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("ai")>();
+  return { ...actual, generateText: mockGenerateText };
+});
+
+const { generateGroundedCopy } = await import("../grounded-copy");
+
+const schema = z.object({ title: z.string(), description: z.string() });
+const fallback = { title: "Static title", description: "Static description" };
+
+const generated = {
+  title: "Bounce rate at 6.0%",
+  description: "This is above the 5% threshold.",
+};
+
+describe("generateGroundedCopy model request", () => {
+  beforeEach(() => {
+    mockGetAIModel.mockReset();
+    mockGenerateText.mockReset();
+    mockGetAIModel.mockResolvedValue({
+      model: { modelId: "claude-haiku-5.5" },
+      modelId: "claude-haiku-5.5",
+      providerOptions: undefined,
+    });
+    mockGenerateText.mockResolvedValue({
+      output: generated,
+      usage: { inputTokens: 10, outputTokens: 20, totalTokens: 30 },
+    });
+  });
+
+  it("requests the model without reasoning so thinking cannot consume the output cap", async () => {
+    await generateGroundedCopy({
+      schema,
+      system: "You write short dashboard copy.",
+      facts: { current: 6 },
+      fallback,
+    });
+
+    expect(mockGetAIModel).toHaveBeenCalledTimes(1);
+    const [request] = mockGetAIModel.mock.calls[0];
+    // `maxOutputTokens` is 300; reasoning tokens count against it and
+    // truncated the JSON, so the request must not enable reasoning.
+    expect(request).toEqual({ model: "claude-haiku-5.5" });
+    expect(request).not.toHaveProperty("reasoning");
+  });
+
+  it("returns the parsed copy when the model answers within the cap", async () => {
+    const result = await generateGroundedCopy({
+      schema,
+      system: "You write short dashboard copy.",
+      facts: { current: 6 },
+      fallback,
+    });
+
+    expect(result).toEqual({ value: generated, generated: true });
+  });
+});

apps/web/src/lib/ai/grounded-copy.ts

@@ -36,8 +36,15 @@ export async function generateGroundedCopy<T>(args: {
   try {
     // A two-line rephrase of established facts: the smallest model is enough,
     // and providers that cannot serve it fall back to their default.
+    //
+    // No `reasoning` here. Thinking tokens count against `maxOutputTokens`,
+    // which is tuned for the two-line answer below, so enabling thinking made
+    // the model spend the budget before the JSON was complete. Every response
+    // came back truncated (finish_reason "length"), structured-output parsing
+    // threw, and every call fell back to the static copy. The copy routes that
+    // do want reasoning pair it with a far larger cap.
     const { model, modelId, providerOptions } = await getAIModel(
-      { model: "claude-haiku-5.5", reasoning: { effort: "low" } },
+      { model: "claude-haiku-5.5" },
       aiEnv()
     );