fix: disable ollama qwen thinking when reasoning off

Co-authored-by: Bee <68532117+abeatrix@users.noreply.github.com>

sdk/packages/llms/src/providers/routing/provider-option-rules.ts

@@ -38,6 +38,11 @@ function isDeepSeekFamily(input: ProviderOptionMatchInput): boolean {
 	return !!input.modelFamily?.trim().toLowerCase().includes("deepseek");
 }
 
+function isQwen3ModelId(input: ProviderOptionMatchInput): boolean {
+	const normalizedModelId = input.request.modelId.toLowerCase();
+	return normalizedModelId.includes("qwen3");
+}
+
 function resolveFamilyThinkingType(
 	input: ProviderOptionMatchInput,
 	defaultWhenUnset: "enabled" | "disabled" | undefined,
@@ -169,6 +174,32 @@ const openRouterReasoningRule: ProviderOptionRule = {
 		),
 };
 
+const ollamaQwen3ReasoningDisabledRule: ProviderOptionRule = {
+	id: "provider.ollama.qwen3.disable-reasoning",
+	phase: "provider-reasoning",
+	description:
+		"Ollama OpenAI-compatible Qwen3 models need reasoning_effort=none to disable default thinking.",
+	applies: (input) =>
+		input.request.providerId === "ollama" &&
+		isQwen3ModelId(input) &&
+		input.request.reasoning?.enabled === false,
+	suppresses: { genericThinking: true },
+	build: (input) => {
+		const bucketOptions = {
+			reasoningEffort: "none",
+			reasoning: { effort: "none" },
+		};
+		return {
+			...buildProviderAndAliasPatch({
+				providerId: input.request.providerId,
+				providerOptionsKey: input.providerOptionsKey,
+				bucketOptions,
+			}),
+			openaiCompatible: bucketOptions,
+		};
+	},
+};
+
 const geminiThinkingRule: ProviderOptionRule = {
 	id: "provider.google-gemini.thinking-config",
 	phase: "provider",
@@ -304,6 +335,7 @@ export const PROVIDER_OPTION_RULES: ReadonlyArray<ProviderOptionRule> = [
 	genericProviderFanoutRule,
 	clineGatewayReasoningRule,
 	openRouterReasoningRule,
+	ollamaQwen3ReasoningDisabledRule,
 	geminiThinkingRule,
 	clineReasoningDisabledThinkingRule,
 	kimiK26ThinkingRule,

sdk/packages/llms/src/providers/routing/provider-options.test.ts

@@ -735,6 +735,52 @@ describe("composeAiSdkProviderOptions: family/provider thinking patches", () =>
 				},
 			],
 		},
+		{
+			name: "ollama Qwen3 modelId reasoning.enabled=false -> effort none in provider buckets",
+			request: {
+				providerId: "ollama",
+				modelId: "qwen3-coder:30b",
+				reasoning: { enabled: false },
+			},
+			context: { family: "qwen" },
+			expect: [
+				{
+					bucket: "ollama",
+					has: {
+						reasoningEffort: "none",
+						reasoning: { effort: "none" },
+					},
+					lacks: ["thinking"],
+				},
+				{
+					bucket: "openaiCompatible",
+					has: {
+						reasoningEffort: "none",
+						reasoning: { effort: "none" },
+					},
+					lacks: ["thinking"],
+				},
+			],
+		},
+		{
+			name: "ollama non-Qwen3 modelId reasoning.enabled=false -> no effort-none override",
+			request: {
+				providerId: "ollama",
+				modelId: "llama3.1",
+				reasoning: { enabled: false },
+			},
+			context: { family: "llama" },
+			expect: [
+				{
+					bucket: "ollama",
+					lacks: ["reasoningEffort", "reasoning", "thinking"],
+				},
+				{
+					bucket: "openaiCompatible",
+					lacks: ["reasoningEffort", "reasoning", "thinking"],
+				},
+			],
+		},
 		{
 			name: "cline non-K2.6 Moonshot Kimi reasoning.enabled=false -> thinking.type=disabled",
 			request: {