fix(ai-gateway-provider): support Azure OpenAI v1 responses routing

.changeset/funny-rivers-protect.md

@@ -0,0 +1,8 @@
+---
+"ai-gateway-provider": patch
+"workers-ai-provider": patch
+---
+
+Support Azure OpenAI v1 Responses API URLs when routing wrapped models through AI Gateway.
+
+`createAzure(...).responses(...)` with `useDeploymentBasedUrls: false` generates `https://{resource}.openai.azure.com/openai/v1/responses?...` URLs, which the shared provider matcher did not recognize, so wrapped Azure responses models failed with `provider "azure.responses" is currently not supported`. These URLs now route through the `azure-openai` gateway provider as `{resource}/openai/responses...` (Azure's non-deployment Responses route). Deployment-based Azure routing is unchanged, and the match is deliberately scoped to `responses` — other v1 paths (chat, embeddings) have no non-deployment Azure route to map to. Applies to both `ai-gateway-provider`'s wrapped-model routing and `workers-ai-provider`'s `createGatewayProvider` URL detection, which share the matcher.

packages/ai-gateway-provider/test/endpoint.test.ts

@@ -52,6 +52,11 @@ const testCases = [
 		name: "azure-openai",
 		url: "https://myresource.openai.azure.com/openai/deployments/mydeployment/chat/completions?api-version=2024-02-15-preview",
 	},
+	{
+		expected: "myresource/openai/responses?api-version=2024-02-15-preview",
+		name: "azure-openai",
+		url: "https://myresource.openai.azure.com/openai/v1/responses?api-version=2024-02-15-preview",
+	},
 	{
 		expected: "v1/chat/completions",
 		name: "openrouter",
@@ -73,6 +78,15 @@ describe("ProvidersConfigs endpoint parsing", () => {
 			expect(result).toBe(testCase.expected);
 		});
 	}
+
+	// Azure's v1 surface is only routable for the Responses API (the sole path
+	// with a non-deployment Azure route); other v1 paths must stay unmatched
+	// rather than route to an endpoint that 404s at Azure.
+	it("does not match Azure v1 paths other than responses", () => {
+		const url =
+			"https://myresource.openai.azure.com/openai/v1/chat/completions?api-version=2024-02-15-preview";
+		expect(providers.find((p) => p.regex.test(url))).toBeUndefined();
+	});
 });
 
 describe("Provider auth header selection", () => {

packages/ai-gateway-provider/test/request-shaping.test.ts

@@ -3,6 +3,7 @@ import { http, HttpResponse } from "msw";
 import { setupServer } from "msw/node";
 import { afterAll, afterEach, beforeAll, describe, expect, it } from "vitest";
 import { CF_TEMP_TOKEN } from "../src/auth";
+import { createAzure } from "../src/providers/azure";
 import { createOpenAI } from "../src/providers/openai";
 import { createAiGateway } from "../src";
 
@@ -113,6 +114,46 @@ describe("REST request shaping", () => {
 		expect(captured.headers["cf-aig-byok-alias"]).toBe("alias-1");
 		expect(captured.headers["cf-aig-zdr"]).toBe("true");
 	});
+
+	it("routes Azure responses API v1 URLs through the azure-openai provider entry", async () => {
+		server.use(
+			http.post(GATEWAY_URL, async ({ request }) => {
+				captured.headers = Object.fromEntries(request.headers);
+				captured.body = await request.json();
+				return HttpResponse.json(openAiResponse());
+			}),
+		);
+
+		const aigateway = createAiGateway({
+			accountId: TEST_ACCOUNT_ID,
+			apiKey: TEST_API_KEY,
+			gateway: TEST_GATEWAY,
+		});
+		const azure = createAzure({
+			resourceName: "myresource",
+			apiKey: "azure-api-key",
+			apiVersion: "2024-02-15-preview",
+			useDeploymentBasedUrls: false,
+		});
+
+		await generateText({
+			model: aigateway(azure.responses("gpt-5.1")),
+			prompt: "hi",
+			maxRetries: 0,
+		});
+
+		const body = captured.body as Array<{
+			provider: string;
+			endpoint: string;
+			query: { model?: string };
+		}>;
+		expect(body).toHaveLength(1);
+		expect(body[0]?.provider).toBe("azure-openai");
+		expect(body[0]?.endpoint).toBe(
+			"myresource/openai/responses?api-version=2024-02-15-preview",
+		);
+		expect(body[0]?.query.model).toBe("gpt-5.1");
+	});
 });
 
 describe("Binding request shaping", () => {

packages/gateway-core/src/gateway-providers.ts

@@ -134,16 +134,29 @@ function bedrockTransform(url: string): string {
 	return `bedrock-runtime/${region}/${rest}`;
 }
 
-// Azure's URL carries the resource + deployment, so it needs a bespoke transform
-// (mirrors ai-gateway-provider). Only used for bring-your-own-provider detection.
+// Azure OpenAI supports both deployment-based URLs
+// (`/openai/deployments/{deployment}/...`) and the v1 surface
+// (`/openai/v1/...`). AI Gateway expects the former as
+// `{resource}/{deployment}/{rest}`. On the v1 surface only the Responses API is
+// routable: Azure exposes a non-deployment `/openai/responses` route, so the
+// gateway endpoint is `{resource}/openai/responses...`. Other v1 paths (chat,
+// embeddings) have no such non-deployment route and stay unmatched rather than
+// being routed to an endpoint that 404s at Azure.
+// Only used for bring-your-own-provider detection.
 const AZURE_HOST =
-	/^https:\/\/(?<resource>[^.]+)\.openai\.azure\.com\/openai\/deployments\/(?<deployment>[^/]+)\/(?<rest>.*)$/;
+	/^https:\/\/(?<resource>[^.]+)\.openai\.azure\.com\/openai\/(?:(?:deployments\/(?<deployment>[^/]+)\/(?<rest>.*))|(?:v1\/(?<v1rest>responses.*)))$/;
 function azureTransform(url: string): string {
 	const m = url.match(AZURE_HOST);
 	if (!m?.groups) return url;
-	const { resource, deployment, rest } = m.groups;
-	if (!resource || !deployment || !rest) return url;
-	return `${resource}/${deployment}/${rest}`;
+	const { resource, deployment, rest, v1rest } = m.groups;
+	if (!resource) return url;
+	if (deployment && rest) {
+		return `${resource}/${deployment}/${rest}`;
+	}
+	if (v1rest) {
+		return `${resource}/openai/${v1rest}`;
+	}
+	return url;
 }
 
 /**