fix(models): reasoning_effort for azure grok-4.3

packages/actions/src/prepare-request-body.spec.ts

@@ -1103,6 +1103,57 @@ describe("prepareRequestBody - Moonshot thinking", () => {
 	});
 });
 
+describe("prepareRequestBody - xAI reasoning_effort", () => {
+	async function prepareXai(
+		provider: Parameters<typeof prepareRequestBody>[0],
+		model: string,
+		externalId: string,
+		effort: Parameters<typeof prepareRequestBody>[14],
+	) {
+		return (await prepareRequestBody(
+			provider,
+			model,
+			null,
+			externalId,
+			[{ role: "user", content: "Hello!" }],
+			false, // stream
+			undefined, // temperature
+			undefined, // max_tokens
+			undefined, // top_p
+			undefined, // frequency_penalty
+			undefined, // presence_penalty
+			undefined, // response_format
+			undefined, // tools
+			undefined, // tool_choice
+			effort,
+			true, // supportsReasoning
+		)) as unknown as Record<string, unknown>;
+	}
+
+	test("forwards reasoning_effort for azure-ai-foundry/grok-4.3", async () => {
+		const requestBody = await prepareXai(
+			"azure-ai-foundry",
+			"grok-4-3",
+			"grok-4.3",
+			"high",
+		);
+		expect(requestBody.reasoning_effort).toBe("high");
+	});
+
+	test.each([
+		["xai", "grok-4", "grok-4"],
+		["xai", "grok-4-20-beta-0309-reasoning", "grok-4.20-beta-0309-reasoning"],
+		["vertex-openai", "grok-4-20-reasoning", "grok-4.20-reasoning"],
+		["xai", "grok-build-0-1", "grok-build-0.1"],
+	] as const)(
+		"drops reasoning_effort for %s/%s, which does not declare it",
+		async (provider, model, externalId) => {
+			const requestBody = await prepareXai(provider, model, externalId, "high");
+			expect(requestBody.reasoning_effort).toBeUndefined();
+		},
+	);
+});
+
 describe("prepareRequestBody - Alibaba thinking", () => {
 	async function prepare(options: {
 		model: string;
@@ -4304,7 +4355,7 @@ describe("prepareRequestBody - max_tokens forwarding", () => {
 			expect(requestBody.max_tokens).toBe(8192);
 			expect(requestBody.max_completion_tokens).toBeUndefined();
 			expect(requestBody.top_p).toBe(0.9);
-			expect(requestBody.reasoning_effort).toBeUndefined();
+			expect(requestBody.reasoning_effort).toBe("medium");
 			expect(requestBody.inferenceConfig).toBeUndefined();
 		});
 	});

packages/models/src/model-metadata.spec.ts

@@ -79,4 +79,25 @@ describe("model metadata", () => {
 
 		expect(offenders).toEqual([]);
 	});
+
+	// xAI accepts `reasoning_effort` on some deployments and rejects it on
+	// others, and the gateway forwards the value verbatim. A mapping that
+	// advertises the parameter has therefore been verified against its
+	// deployment, and that verification is what `reasoningEfforts` records —
+	// without it neither the gateway nor a client knows which tiers are safe.
+	it("declares reasoningEfforts for every xAI mapping that accepts reasoning_effort", () => {
+		const offenders = (models as readonly ModelDefinition[])
+			.filter((model) => model.family === "xai")
+			.flatMap((model) =>
+				(model.providers as readonly ProviderModelMapping[])
+					.filter(
+						(provider) =>
+							provider.supportedParameters?.includes("reasoning_effort") &&
+							!provider.reasoningEfforts?.length,
+					)
+					.map((provider) => `${provider.providerId}/${model.id}`),
+			);
+
+		expect(offenders).toEqual([]);
+	});
 });

packages/models/src/models/xai.ts

@@ -1,5 +1,11 @@
 import type { ModelDefinition } from "@/models.js";
 
+// `reasoning_effort` is deliberately absent: xAI accepts it only on some
+// deployments (grok-4.3 and grok-4.5 do, grok-4 does not), and the gateway
+// forwards it verbatim, so a mapping that advertises it on a deployment that
+// rejects it turns working requests into upstream 400s. A mapping only leaves
+// this list once its deployment is verified live, and then it also declares
+// `reasoningEfforts` so clients can discover the tiers it accepts.
 const xaiSupportedParamsNoFreqPresence = [
 	"temperature",
 	"max_tokens",
@@ -761,6 +767,7 @@ export const xaiModels = [
 				streaming: true,
 				vision: false,
 				reasoning: true,
+				reasoningEfforts: ["none", "low", "medium", "high"],
 				reasoningOutput: "omit",
 				tools: true,
 				jsonOutput: false,
@@ -770,6 +777,7 @@ export const xaiModels = [
 					"top_p",
 					"tools",
 					"tool_choice",
+					"reasoning_effort",
 				],
 			},
 		],