fix(ai): handle explicit thinking off across providers closes #2490

packages/ai/src/providers/google-vertex.ts

@@ -436,6 +436,8 @@ function buildParams(
 			thinkingConfig.thinkingBudget = options.thinking.budgetTokens;
 		}
 		config.thinkingConfig = thinkingConfig;
+	} else if (model.reasoning && options.thinking && !options.thinking.enabled) {
+		config.thinkingConfig = getDisabledThinkingConfig(model);
 	}
 
 	if (options.signal) {
@@ -464,6 +466,22 @@ function isGemini3FlashModel(model: Model<"google-generative-ai">): boolean {
 	return /gemini-3(?:\.\d+)?-flash/.test(model.id.toLowerCase());
 }
 
+function getDisabledThinkingConfig(model: Model<"google-vertex">): ThinkingConfig {
+	// Google docs: Gemini 3.1 Pro cannot disable thinking, and Gemini 3 Flash / Flash-Lite
+	// do not support full thinking-off either. For Gemini 3 models, use the lowest supported
+	// thinkingLevel without includeThoughts so hidden thinking remains invisible to pi.
+	const geminiModel = model as unknown as Model<"google-generative-ai">;
+	if (isGemini3ProModel(geminiModel)) {
+		return { thinkingLevel: ThinkingLevel.LOW };
+	}
+	if (isGemini3FlashModel(geminiModel)) {
+		return { thinkingLevel: ThinkingLevel.MINIMAL };
+	}
+
+	// Gemini 2.x supports disabling via thinkingBudget = 0.
+	return { thinkingBudget: 0 };
+}
+
 function getGemini3ThinkingLevel(
 	effort: ClampedThinkingLevel,
 	model: Model<"google-generative-ai">,