fix(agent): route gateway-served qwen models through Veryfront Cloud

src/agent/runtime/model-resolution.test.ts

@@ -10,6 +10,7 @@ import { deleteEnv, setEnv } from "#veryfront/compat/process.ts";
 import { afterEach, describe, it } from "#veryfront/testing/bdd.ts";
 import {
   resolveVeryfrontCloudModelId,
+  VERYFRONT_CLOUD_CATALOG_PROVIDER_NAMES,
   VERYFRONT_CLOUD_CHAT_MODELS,
 } from "#veryfront/provider/veryfront-cloud/model-catalog.ts";
 import {
@@ -394,6 +395,53 @@ describe("agent/runtime/model-resolution", () => {
     );
   });
 
+  it("routes every catalog provider through veryfront-cloud when only hosted bootstrap is available", () => {
+    setEnv("VERYFRONT_API_TOKEN", "vf_test_runtime");
+    setEnv("VERYFRONT_PROJECT_SLUG", "demo-project");
+
+    for (const provider of VERYFRONT_CLOUD_CATALOG_PROVIDER_NAMES) {
+      // Mistral model IDs are gated by the catalog, so it needs a listed one.
+      const modelId = provider === "mistral" ? "mistral-small-2503" : "model-x";
+      assertEquals(
+        resolveRuntimeModel(`${provider}/${modelId}`),
+        `veryfront-cloud/${provider}/${modelId}`,
+        provider,
+      );
+    }
+  });
+
+  it("routes every vendor the gateway catalog serves through veryfront-cloud (#1913)", () => {
+    // The vendors GET /ai/models lists. Qwen is served before the catalog
+    // snapshot in this package names it; a vendor missing here fails hosted
+    // runs with `Model provider "<vendor>" not registered`.
+    setEnv("VERYFRONT_API_TOKEN", "vf_test_runtime");
+    setEnv("VERYFRONT_PROJECT_SLUG", "demo-project");
+
+    assertEquals(
+      [
+        "anthropic/claude-sonnet-4-6",
+        "openai/gpt-5-nano",
+        "google/gemini-3.5-flash",
+        "mistral/mistral-small-2503",
+        "deepseek/deepseek-v4-flash",
+        "qwen/qwen3.8-27b",
+      ].map((model) => resolveRuntimeModel(model)),
+      [
+        "veryfront-cloud/anthropic/claude-sonnet-4-6",
+        "veryfront-cloud/openai/gpt-5-nano",
+        "veryfront-cloud/google/gemini-3.5-flash",
+        "veryfront-cloud/mistral/mistral-small-2503",
+        "veryfront-cloud/deepseek/deepseek-v4-flash",
+        "veryfront-cloud/qwen/qwen3.8-27b",
+      ],
+    );
+  });
+
+  it("keeps a gateway-only provider unrouted without hosted bootstrap", () => {
+    setEnv("OPENAI_API_KEY", "sk-test");
+    assertEquals(resolveRuntimeModel("qwen/qwen3.8-27b"), "qwen/qwen3.8-27b");
+  });
+
   it("routes catalog Gemini, Mistral, and Kimi models through veryfront-cloud when only hosted bootstrap is available", () => {
     setEnv("VERYFRONT_API_TOKEN", "vf_test_runtime");
     setEnv("VERYFRONT_PROJECT_SLUG", "demo-project");

src/agent/runtime/model-resolution.ts

@@ -8,6 +8,7 @@ import {
   createRetiredVeryfrontCloudModelError,
   findVeryfrontCloudModelByModelId,
   isRetiredVeryfrontCloudModelId,
+  VERYFRONT_CLOUD_CATALOG_PROVIDER_NAMES,
 } from "#veryfront/provider/veryfront-cloud/model-catalog.ts";
 import { DEFAULT_MODEL_CREDENTIAL_MISMATCH, NOT_SUPPORTED } from "#veryfront/errors";
 import {
@@ -21,14 +22,16 @@ import { getModelRuntimeProvider } from "#veryfront/provider/runtime-inspection.
 export const AUTO_AGENT_MODEL = "auto";
 export const DEFAULT_AGENT_MODEL = "openai/gpt-5-nano";
 
-const HOSTED_PROVIDER_NAMES = new Set([
-  "deepseek",
-  "anthropic",
-  "google",
-  "google-ai-studio",
-  "mistral",
-  "moonshotai",
-  "openai",
+/**
+ * Providers the gateway serves that the catalog snapshot shipped in this
+ * package does not list yet. Each routes on the default surface, which the
+ * gateway serves at the vendor-neutral `/ai/v1`. Drop an entry once
+ * `deno task generate:model-catalog` adds its provider to the snapshot.
+ */
+const GATEWAY_PROVIDERS_AHEAD_OF_CATALOG = ["qwen"] as const;
+const HOSTED_PROVIDER_NAMES: ReadonlySet<string> = new Set([
+  ...VERYFRONT_CLOUD_CATALOG_PROVIDER_NAMES,
+  ...GATEWAY_PROVIDERS_AHEAD_OF_CATALOG,
 ]);
 const DIRECT_CREDENTIAL_PROVIDER_ALIASES = new Map<string, string>([
   ["google-ai-studio", "google"],

src/provider/veryfront-cloud/model-catalog.ts

@@ -415,6 +415,20 @@ export const VERYFRONT_CLOUD_CHAT_MODELS: readonly VeryfrontCloudChatModel[] = O
   }),
 );
 
+/**
+ * Every provider name the catalog data routes: accepted aliases, providers
+ * with a routing row, and providers of a listed chat model. Runtime model
+ * resolution sends `<provider>/<model>` for these through the gateway when no
+ * direct provider credential applies.
+ */
+export const VERYFRONT_CLOUD_CATALOG_PROVIDER_NAMES: readonly string[] = Object.freeze([
+  ...new Set<string>([
+    ...VERYFRONT_CLOUD_PROVIDER_ALIASES.map(([alias]) => alias),
+    ...VERYFRONT_CLOUD_PROVIDER_ROUTING.map(([provider]) => provider),
+    ...VERYFRONT_CLOUD_CHAT_MODELS.map((model) => model.provider),
+  ]),
+]);
+
 const defaultVeryfrontCloudChatModel = VERYFRONT_CLOUD_CHAT_MODELS.find(
   (model) => model.id === DEFAULT_VERYFRONT_CLOUD_MODEL_ID,
 );

src/provider/veryfront-cloud/provider.test.ts

@@ -1330,6 +1330,44 @@ describe("provider/veryfront-cloud", () => {
     assertEquals(result.text, "Hello");
   });
 
+  it("sends a hosted qwen/qwen3.8-27b run to the vendor-neutral gateway route (#1913)", async () => {
+    setCloudBootstrap();
+    const encoder = new TextEncoder();
+    let captured: { url: string; model: unknown } | undefined;
+
+    installMockFetch(
+      (async (input: URL | Request | string, init?: RequestInit) => {
+        const request = new Request(input, init);
+        captured = { url: request.url, model: JSON.parse(await request.text()).model };
+
+        return new Response(
+          new ReadableStream({
+            start(controller) {
+              controller.enqueue(
+                encoder.encode('data: {"choices":[{"delta":{"content":"Hello"}}]}\n\n'),
+              );
+              controller.enqueue(
+                encoder.encode('data: {"choices":[{"finish_reason":"stop"}]}\n\n'),
+              );
+              controller.enqueue(encoder.encode("data: [DONE]\n\n"));
+              controller.close();
+            },
+          }),
+          { status: 200, headers: { "content-type": "text/event-stream" } },
+        );
+      }) as typeof fetch,
+    );
+
+    const assistant = agent({ model: "qwen/qwen3.8-27b", system: "You are concise." });
+    const result = await assistant.generate({ input: "Hi" });
+
+    assertEquals(captured, {
+      url: "https://api.veryfront.com/ai/v1/chat/completions",
+      model: "qwen/qwen3.8-27b",
+    });
+    assertEquals(result.text, "Hello");
+  });
+
   it("keeps an unlisted provider on chat completions for a reasoning-style model id", async () => {
     // "gpt-5.4" is a reasoning-style ID. Only the provider that implements the
     // OpenAI surface natively serves /responses, so an unlisted provider must