feat(xai): expose grok-4.20-multi-agent on the Responses lane

scripts/model-metadata.source.json

@@ -81148,6 +81148,35 @@
 				"maxLevel": "high"
 			}
 		},
+		"grok-4.20-multi-agent-0309": {
+			"id": "grok-4.20-multi-agent-0309",
+			"name": "Grok 4.20 Multi-Agent (0309)",
+			"api": "openai-responses",
+			"provider": "xai",
+			"baseUrl": "https://api.x.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 1.25,
+				"output": 2.5,
+				"cacheRead": 0.2,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1000000,
+			"maxTokens": 30000,
+			"compat": {
+				"supportsImageDetailOriginal": false,
+				"includeEncryptedReasoning": false
+			},
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "minimal",
+				"maxLevel": "high"
+			}
+		},
 		"grok-4.20-multi-agent-beta-latest": {
 			"id": "grok-4.20-multi-agent-beta-latest",
 			"name": "Grok 4.20 Multi-Agent Beta",

src/codex/catalog/provider-fetch.ts

@@ -961,6 +961,7 @@ export const CALLABLE_CONFIGURED_COMPATIBILITY_MODELS: Readonly<Record<string, R
   ]),
   xai: new Set([
     "grok-4.3",
+    "grok-4.20-multi-agent-0309",
     "grok-4.20-0309-reasoning",
     "grok-4.20-0309-non-reasoning",
     "grok-build-0.1",
@@ -1564,6 +1565,9 @@ export async function fetchProviderModels(
 
 export function shouldExposeProviderModel(providerName: string, modelId: string): boolean {
   if (providerName === "opencode-free") return modelId === "big-pickle" || modelId.endsWith("-free");
+  // xAI /models advertises both the dated deployment and this floating alias.
+  // Keep only grok-4.20-multi-agent-0309; the alias is the same server-side id.
+  if (providerName === "xai" && modelId === "grok-4.20-multi-agent-beta-latest") return false;
   return true;
 }
 

src/providers/registry.ts

@@ -454,6 +454,7 @@ const XAI_MODELS = [
   "grok-4.6",
   "grok-4.5",
   "grok-4.3",
+  "grok-4.20-multi-agent-0309",
   "grok-4.20-0309-reasoning",
   "grok-4.20-0309-non-reasoning",
   "grok-build-0.1",
@@ -1148,8 +1149,9 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     // 260709 refresh: lineup + metadata from official docs.x.ai (grok-4.5 announced 07-08);
     // grok-composer-2.5-fast kept as account-verified (absent from public docs). Evidence:
     // devlog/model_update/260709_model_refresh/001_xai_lineup.md.
-    // grok-4.20-multi-agent-0309 is intentionally absent: the OAuth chat-completions
-    // transport returns 400 ("Multi Agent requests are not allowed on chat completions").
+    // 260823: grok-4.20-multi-agent-0309 still returns 400 on Chat Completions, but works
+    // on Responses. The server reports this dated id for both it and the floating
+    // grok-4.20-multi-agent-beta-latest alias, so expose only the dated deployment id.
     // 260813: grok-4.6 added per docs.x.ai/developers/grok-4-6. Context/vision still match
     // grok-4.5; the reasoning ladder does not — 4.6 adds the documented xhigh rung.
     models: XAI_MODELS,
@@ -1165,9 +1167,11 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
     // than the seeded ones do.
     supportsVerbosity: false,
     defaultModel: "grok-4.5",
-    // Keep Codex Responses callers on the compatibility Chat wire until xAI can replay
-    // opaque reasoning continuation and compaction state across later turns. The scoped
-    // declaration also keeps caller-owned service tiers off the OAuth subscription route.
+    // Keep 4.6/4.5 Responses callers on the compatibility Chat wire until xAI can replay
+    // opaque reasoning continuation and compaction state across later turns. Multi-agent has
+    // no Chat wire, so Responses callers use its only working wire under both auth modes.
+    // Caller-owned service tiers stay off the unclassified OAuth subscription route; key-auth
+    // Fast remains proxy-owned and is still selected through keyAuthServiceTier above.
     modelWireDefaults: {
       "grok-4.6": {
         wire: "openai-chat",
@@ -1181,6 +1185,19 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
         authModes: ["oauth"],
         forwardCallerServiceTier: false,
       },
+      "grok-4.20-multi-agent-0309": {
+        // Even at high effort it emits no reasoning-summary deltas or encrypted replay
+        // material. Do not encode that as modelSupportsReasoningSummaries:false: through
+        // Codex #1100 that suppresses the entire reasoning object, including the effort
+        // that controls this model's agent count. An empty summary pane is harmless.
+        // Chat Completions returns 400 for this model, so every inbound uses Responses —
+        // `anthropic` included. Omitting it left providerModelWireDefault returning undefined
+        // for the Claude Messages lane, so resolveWireProtocolOverride kept xAI's provider-wide
+        // openai-chat adapter and sent this model to the wire it 400s on.
+        wire: "openai-responses",
+        inbound: ["responses", "chat", "anthropic"],
+        forwardCallerServiceTier: false,
+      },
     },
     // Vision lineup per docs.x.ai model-capabilities/images/understanding: the grok-4.x chat
     // models accept image input (JPEG/PNG, URL or base64). Without this the catalog leaves
@@ -1192,6 +1209,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       "grok-4.6": ["text", "image"],
       "grok-4.5": ["text", "image"],
       "grok-4.3": ["text", "image"],
+      "grok-4.20-multi-agent-0309": ["text", "image"],
       "grok-4.20-0309-reasoning": ["text", "image"],
       "grok-4.20-0309-non-reasoning": ["text", "image"],
     },
@@ -1210,6 +1228,7 @@ export const PROVIDER_REGISTRY: readonly ProviderRegistryEntry[] = [
       "grok-4.6": 500_000,
       "grok-4.5": 500_000,
       "grok-4.3": 1_000_000,
+      "grok-4.20-multi-agent-0309": 1_000_000,
       "grok-4.20-0309-reasoning": 1_000_000,
       "grok-4.20-0309-non-reasoning": 1_000_000,
       "grok-build-0.1": 256_000,

tests/codex-catalog.test.ts

@@ -3732,13 +3732,28 @@ describe("Codex catalog routed normalization", () => {
         "kimi/kimi-k2.7-code-highspeed",
         "xai/grok-4.20-0309-non-reasoning",
         "xai/grok-4.20-0309-reasoning",
+        "xai/grok-4.20-multi-agent-0309",
         "xai/grok-4.3",
         "xai/grok-4.5",
         "xai/grok-build-0.1",
         "xai/grok-composer-2.5-fast",
       ]);
       expect(models.find(model => model.provider === "kimi" && model.id === "k3[1m]")?.contextWindow).toBe(1_048_576);
-      expect(models.some(model => model.id === "grok-4.20-multi-agent-0309")).toBe(false);
+      expect(models.find(model => model.provider === "xai" && model.id === "grok-4.20-multi-agent-0309"))
+        .toMatchObject({
+          contextWindow: 1_000_000,
+          inputModalities: ["text", "image"],
+        });
+      expect(models.find(model => model.provider === "xai" && model.id === "grok-4.20-multi-agent-0309")?.supportsReasoningSummaries)
+        .toBeUndefined();
+      expect(getModelMetadata("xai", "grok-4.20-multi-agent-0309")).toMatchObject({
+        contextWindow: 1_000_000,
+        maxTokens: 30_000,
+        input: ["text", "image"],
+        reasoning: true,
+        cost: { input: 1.25, output: 2.5, cacheRead: 0.2, cacheWrite: 0 },
+      });
+      expect(getModelMetadata("xai", "grok-4.20-multi-agent-0309")).not.toHaveProperty("supportsReasoningSummaries");
       expect(models.some(model => model.id === "configured-ghost")).toBe(false);
       expect(warning.mock.calls.flat().join(" ")).not.toContain("omitted configured model ids");
     } finally {

tests/provider-live-models.test.ts

@@ -41,6 +41,7 @@ afterEach(() => {
   clearModelCache(PROVIDER);
   clearModelCache(HY3_PROVIDER);
   clearModelCache(HY3_CONTROL_PROVIDER);
+  clearModelCache("xai");
 });
 
 describe("live provider model discovery (authority + fallback)", () => {
@@ -79,6 +80,34 @@ describe("live provider model discovery (authority + fallback)", () => {
     }
   });
 
+  test("xAI live discovery hides the multi-agent beta alias and keeps the dated deployment", async () => {
+    globalThis.fetch = (async () => new Response(JSON.stringify({
+      data: [
+        { id: "grok-4.20-multi-agent-0309" },
+        { id: "grok-4.20-multi-agent-beta-latest" },
+        { id: "grok-4.6" },
+      ],
+    }), { status: 200, headers: { "content-type": "application/json" } })) as typeof fetch;
+
+    const models = await gatherRoutedModels({
+      providers: {
+        xai: withTestFetch({
+          baseUrl: "https://api.x.ai/v1",
+          adapter: "openai-chat",
+          authMode: "key",
+          apiKey: "sk-test",
+          liveModels: true,
+          models: ["grok-4.6"],
+        }),
+      },
+    } as unknown as OcxConfig);
+    const ids = models.filter(model => model.provider === "xai").map(model => model.id);
+
+    expect(ids).toContain("grok-4.20-multi-agent-0309");
+    expect(ids).toContain("grok-4.6");
+    expect(ids).not.toContain("grok-4.20-multi-agent-beta-latest");
+  });
+
   test("HY3 compatibility guard hides only opencode-go/hy3-preview from live discovery", async () => {
     globalThis.fetch = (async (url: string | URL | Request) => {
       const isOpenCodeGo = String(url).startsWith("https://opencode-go.test/");

tests/xai-transport.test.ts

@@ -11,6 +11,8 @@ import {
   XAI_GROK_CLIENT_VERSION,
 } from "../src/providers/xai-transport";
 import { getProviderRegistryEntry } from "../src/providers/registry";
+import { XAI_RESPONSES_OPT_IN_MODELS } from "../src/providers/xai-responses-opt-in";
+import { resolveWireProtocolOverride } from "../src/server/adapter-resolve";
 import type { OcxAssistantMessage, OcxParsedRequest, OcxProviderConfig } from "../src/types";
 
 const UUID_V4 = /^[0-9a-f]{8}-[0-9a-f]{4}-4[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i;
@@ -604,15 +606,58 @@ describe("xAI outbound compatibility headers", () => {
 });
 
 describe("xAI reasoning_content cache preservation", () => {
-  test("registry preset replays reasoning_content for grok reasoning models only", () => {
+  test("registry preset exposes multi-agent only on Responses without claiming replay material", () => {
     const entry = getProviderRegistryEntry("xai");
     expect(entry?.preserveReasoningContentModels).toEqual([
       "grok-4.6",
       "grok-4.5",
       "grok-4.3",
       "grok-4.20-0309-reasoning",
     ]);
-    expect(entry?.models).not.toContain("grok-4.20-multi-agent-0309");
+    expect(entry?.models).toContain("grok-4.20-multi-agent-0309");
+    expect(entry?.preserveReasoningContentModels).not.toContain("grok-4.20-multi-agent-0309");
+    expect(entry?.modelSupportsReasoningSummaries?.["grok-4.20-multi-agent-0309"]).toBeUndefined();
+    expect(resolveWireProtocolOverride(
+      "xai",
+      "grok-4.20-multi-agent-0309",
+      provider("oauth"),
+      "responses",
+    ).adapter).toBe("openai-responses");
+    expect(resolveWireProtocolOverride(
+      "xai",
+      "grok-4.20-multi-agent-0309",
+      provider("key"),
+      "responses",
+    ).adapter).toBe("openai-responses");
+    expect(resolveWireProtocolOverride(
+      "xai",
+      "grok-4.20-multi-agent-0309",
+      provider("oauth"),
+      "chat",
+    ).adapter).toBe("openai-responses");
+    expect(resolveWireProtocolOverride(
+      "xai",
+      "grok-4.20-multi-agent-0309",
+      provider("key"),
+      "chat",
+    ).adapter).toBe("openai-responses");
+    // The Claude Messages lane resolves with inbound "anthropic" (src/server/claude-messages.ts).
+    // It is not a spelling of "responses": an inbound missing from the allow-list makes
+    // providerModelWireDefault return undefined, which silently keeps xAI's provider-wide
+    // openai-chat adapter — the one wire this model answers with a 400.
+    expect(resolveWireProtocolOverride(
+      "xai",
+      "grok-4.20-multi-agent-0309",
+      provider("oauth"),
+      "anthropic",
+    ).adapter).toBe("openai-responses");
+    expect(resolveWireProtocolOverride(
+      "xai",
+      "grok-4.20-multi-agent-0309",
+      provider("key"),
+      "anthropic",
+    ).adapter).toBe("openai-responses");
+    expect(XAI_RESPONSES_OPT_IN_MODELS).not.toContain("grok-4.20-multi-agent-0309");
     expect(entry?.models).toContain("grok-build-0.1");
     for (const noReasoning of entry?.noReasoningModels ?? []) {
       expect(entry?.preserveReasoningContentModels).not.toContain(noReasoning);