chore(vendor): import pi 1.3.0

Imported as a linear vendor commit so forge rebases do not replay the upstream root tree.

git-subtree-dir: third_party/pi

git-subtree-split: 71023aec1546d6d677a65669c5665c0927ce0dd0

third_party/pi/packages/ai/test/anthropic-cache-write-1h-cost.test.ts

@@ -1,8 +1,7 @@
 import type Anthropic from "@anthropic-ai/sdk";
 import { describe, expect, it } from "vitest";
 import { stream as streamAnthropic } from "../src/api/anthropic-messages.ts";
-import { getModel } from "../src/compat.ts";
-import type { Context } from "../src/types.ts";
+import { getModel, normalizeContext } from "../src/compat.ts";
 
 function createSseResponse(events: Array<{ event: string; data: string }>): Response {
 	const body = events.map(({ event, data }) => `event: ${event}\ndata: ${data}\n`).join("\n");
@@ -57,7 +56,7 @@ function eventsWithCacheCreation(
 }
 
 // claude-opus-4-8: input 5, cacheWrite (5m) 6.25 per Mtok. 1h write = 2x input = 10.
-const context: Context = { messages: [{ role: "user", content: "hi", timestamp: Date.now() }] };
+const context = normalizeContext({ messages: [{ role: "user", content: "hi", timestamp: Date.now() }] });
 
 describe("Anthropic 1h cache write cost", () => {
 	it("prices the 1h portion at 2x input and the rest at the 5m rate", async () => {
@@ -65,7 +64,9 @@ describe("Anthropic 1h cache write cost", () => {
 		const response = createSseResponse(
 			eventsWithCacheCreation({ ephemeral_5m_input_tokens: 600_000, ephemeral_1h_input_tokens: 400_000 }),
 		);
-		const result = await streamAnthropic(model, context, { client: createFakeAnthropicClient(response) }).result();
+		const result = await streamAnthropic(model, context, {
+			client: createFakeAnthropicClient(response),
+		}).result();
 
 		expect(result.usage.cacheWrite).toBe(1_000_000);
 		expect(result.usage.cacheWrite1h).toBe(400_000);
@@ -73,10 +74,47 @@ describe("Anthropic 1h cache write cost", () => {
 		expect(result.usage.cost.cacheWrite).toBeCloseTo(7.75, 10);
 	});
 
+	// Regression for #9210: Vercel AI Gateway sends cache usage in message_delta, not message_start.
+	it("prices 1h cache writes reported only in message_delta", async () => {
+		const model = getModel("vercel-ai-gateway", "anthropic/claude-haiku-4.5");
+		const response = createSseResponse([
+			{
+				event: "message_start",
+				data: JSON.stringify({
+					type: "message_start",
+					message: { id: "msg_test", usage: { input_tokens: 0, output_tokens: 0 } },
+				}),
+			},
+			{
+				event: "message_delta",
+				data: JSON.stringify({
+					type: "message_delta",
+					delta: { stop_reason: "end_turn" },
+					usage: {
+						input_tokens: 3,
+						output_tokens: 4,
+						cache_creation_input_tokens: 6535,
+						cache_creation: { ephemeral_5m_input_tokens: 0, ephemeral_1h_input_tokens: 6535 },
+					},
+				}),
+			},
+			{ event: "message_stop", data: JSON.stringify({ type: "message_stop" }) },
+		]);
+		const result = await streamAnthropic(model, context, {
+			client: createFakeAnthropicClient(response),
+		}).result();
+
+		expect(result.usage.cacheWrite).toBe(6535);
+		expect(result.usage.cacheWrite1h).toBe(6535);
+		expect(result.usage.cost.cacheWrite).toBeCloseTo((6535 * model.cost.input * 2) / 1_000_000, 10);
+	});
+
 	it("falls back to the 5m rate when no breakdown is reported", async () => {
 		const model = getModel("anthropic", "claude-opus-4-8");
 		const response = createSseResponse(eventsWithCacheCreation(undefined));
-		const result = await streamAnthropic(model, context, { client: createFakeAnthropicClient(response) }).result();
+		const result = await streamAnthropic(model, context, {
+			client: createFakeAnthropicClient(response),
+		}).result();
 
 		expect(result.usage.cacheWrite).toBe(1_000_000);
 		expect(result.usage.cacheWrite1h ?? 0).toBe(0);