LLMs: Gemini: remove 'medium' thinking level support - too early. Fixes #900

src/common/stores/llms/llms.parameters.ts

@@ -208,7 +208,8 @@ export const DModelParameterRegistry = {
     label: 'Thinking Level',
     type: 'enum' as const,
     description: 'Controls internal reasoning depth. Replaces thinking_budget for Gemini 3 models. When unset, the model decides dynamically.',
-    values: ['high', 'medium' /* not present at launch */, 'low' /* default when unset */] as const,
+    // Note: 'medium' and 'minimal' will be available when Gemini 3 Flash launches
+    values: ['high', 'low' /* default when unset */] as const,
     // No initialValue - undefined means 'dynamic', which for Gemini Pro is the same as 'high' (which is the equivalent of 'medium' for OpenAI's effort levels.. somehow)
   } as const,
 

src/modules/aix/server/api/aix.wiretypes.ts

@@ -466,6 +466,8 @@ export namespace AixWire_API {
     vndGeminiMediaResolution: z.enum(['mr_high', 'mr_medium', 'mr_low']).optional(),
     vndGeminiShowThoughts: z.boolean().optional(),
     vndGeminiThinkingBudget: z.number().optional(), // old param
+    // 'medium' will be added when Gemini 3 Flash launches
+    // but we have it below, because we used to have it and we don't want to throw on zod
     vndGeminiThinkingLevel: z.enum(['high', 'medium', 'low']).optional(), // new param
     vndGeminiUrlContext: z.enum(['auto']).optional(),
     // Moonshot

src/modules/aix/server/dispatch/chatGenerate/adapters/gemini.generateContent.ts

@@ -100,7 +100,8 @@ export function aixToGeminiGenerateContent(model: AixAPI_Model, _chatGenerate: A
     // [Gemini 3, 2025-11-18] Thinking Level (replaces thinkingBudget for Gemini 3)
     // CRITICAL: Cannot use both thinkingLevel and thinkingBudget (400 error)
     if (model.vndGeminiThinkingLevel) {
-      thinkingConfig.thinkingLevel = model.vndGeminiThinkingLevel;
+      // FIXME: remove this cast once the 'medium' level is supported upstream
+      thinkingConfig.thinkingLevel = model.vndGeminiThinkingLevel === 'medium' ? 'high' : model.vndGeminiThinkingLevel;
     }
     // [Gemini 2.x] Thinking Budget (0 disables thinking explicitly)
     else if (model.vndGeminiThinkingBudget !== undefined) {

src/modules/aix/server/dispatch/wiretypes/gemini.wiretypes.ts

@@ -569,12 +569,12 @@ export namespace GeminiWire_API_Generate_Content {
       /**
        * [Gemini 3, 2025-11-18] Replaces thinkingBudget for Gemini 3 models.
        * - 'low': Minimizes latency and cost
-       * - 'medium': Balanced (coming soon, not yet supported)
        * - 'high': Maximizes reasoning depth (default when set)
        * - undefined: Dynamic (model decides - which is equivalent to 'high' for now)
+       * Note: 'medium' and 'minimal' will be available when Gemini 3 Flash launches
        * CRITICAL: Cannot use both thinkingLevel and thinkingBudget (400 error)
        */
-      thinkingLevel: z.enum(['low', 'medium', 'high']).optional(),
+      thinkingLevel: z.enum(['low', 'high']).optional(),
     }).optional(),
 
     // Image generation configuration

src/modules/llms/models-modal/LLMParametersEditor.tsx

@@ -111,7 +111,7 @@ const _geminiMediaResolutionOptions = [
 
 const _geminiThinkingLevelOptions = [
   { value: 'high', label: 'High', description: 'Maximum reasoning depth' },
-  { value: 'medium', label: 'Medium', description: 'Balanced reasoning' },
+  // Note: 'medium' and 'minimal' will be available when Gemini 3 Flash launches
   { value: 'low', label: 'Low', description: 'Quick responses (default when unset)' },
   { value: _UNSPECIFIED, label: 'Default', description: 'Model decides automatically (default)' },
 ] as const;