fix(gateway): extract vertex-anthropic streaming token usage

apps/gateway/src/chat/tools/extract-token-usage.spec.ts

@@ -361,6 +361,50 @@ describe("extractTokenUsage", () => {
 		});
 	});
 
+	describe("vertex-anthropic", () => {
+		// Vertex Anthropic streams the same Anthropic Messages wire format, so it
+		// must extract usage identically to the direct anthropic provider rather
+		// than falling through to the OpenAI-shaped default branch.
+		it("extracts cache and reasoning tokens from a streaming message_start", () => {
+			const data = {
+				type: "message_start",
+				message: {
+					usage: {
+						input_tokens: 100,
+						cache_creation_input_tokens: 50,
+						cache_read_input_tokens: 800,
+						output_tokens: 2928,
+						output_tokens_details: { thinking_tokens: 1502 },
+					},
+				},
+			};
+
+			const result = extractTokenUsage(data, "vertex-anthropic");
+
+			expect(result.promptTokens).toBe(950); // 100 + 50 + 800
+			expect(result.cachedTokens).toBe(800);
+			expect(result.cacheCreationTokens).toBe(50);
+			expect(result.completionTokens).toBe(2928);
+			expect(result.reasoningTokens).toBe(1502);
+			expect(result.totalTokens).toBe(3878); // 950 + 2928, reasoning not double-counted
+		});
+
+		it("matches the direct anthropic provider for the same chunk", () => {
+			const data = {
+				usage: {
+					input_tokens: 51,
+					cache_read_input_tokens: 20,
+					output_tokens: 136,
+					reasoning_output_tokens: 31,
+				},
+			};
+
+			expect(extractTokenUsage(data, "vertex-anthropic")).toEqual(
+				extractTokenUsage(data, "anthropic"),
+			);
+		});
+	});
+
 	describe("alibaba", () => {
 		it("extracts prompt_tokens_details.cache_creation_input_tokens into 5m cache write fields", () => {
 			const data = {

apps/gateway/src/chat/tools/extract-token-usage.ts

@@ -167,6 +167,7 @@ export function extractTokenUsage(
 			}
 			break;
 		case "anthropic":
+		case "vertex-anthropic":
 			{
 				const usage = data.message?.usage ?? data.usage;
 				if (!usage) {