fix(gemini): send the lowest thinking level when thinking is 'none' on Gemini 3

Author: waleedlatif1

apps/sim/providers/gemini/core.test.ts

@@ -139,6 +139,20 @@ describe('Vertex Gemini request compatibility', () => {
     })
   })
 
+  /** Omitting thinkingConfig leaves Gemini 3 thinking at its default, so 'none' must send a level. */
+  it.each([
+    ['vertex/gemini-3.6-flash', 'MINIMAL'],
+    ['vertex/gemini-3.8-flash', 'LOW'],
+  ])("sends the lowest thinking level for 'none' on %s", async (model, thinkingLevel) => {
+    const generateContent = vi.fn().mockResolvedValue(textTurn())
+    await run(model, generateContent, { thinkingLevel: 'none' })
+
+    expect(generateContent.mock.calls[0][0].config.thinkingConfig).toEqual({
+      thinkingLevel,
+      includeThoughts: false,
+    })
+  })
+
   it.each(['vertex/gemini-2.5-flash', 'vertex/Custom-Gemini-Model'])(
     'preserves temperature for supported or uncataloged %s',
     async (model) => {

apps/sim/providers/gemini/core.ts

@@ -36,7 +36,7 @@ import {
   mapToThinkingLevel,
   supportsDisablingGemini25Thinking,
 } from '@/providers/google/utils'
-import { getModelCapabilities, isKnownModelId } from '@/providers/models'
+import { getModelCapabilities, getThinkingLevelsForModel, isKnownModelId } from '@/providers/models'
 import { withProviderRetry } from '@/providers/retry'
 import { executeProviderTool } from '@/providers/runtime-context'
 import { createSettledAgentEventStream } from '@/providers/stream-events'
@@ -1010,17 +1010,24 @@ export async function executeGeminiRequest(
     // Gemini 3 and later take thinkingLevel directly; Gemini 2.5-series rejects it and needs thinkingBudget.
     // includeThoughts is required for thought parts to appear; it is requested
     // only on agent-events runs so legacy runs keep the pre-agent-events payload.
-    if (request.thinkingLevel && request.thinkingLevel !== 'none') {
+    const isGemini3 = isGemini3OrLaterModel(model)
+    // Gemini 3 has no off switch and omitting thinkingConfig means the model's default level, so
+    // 'none' sends the lowest declared level (levels are listed lowest-first; MINIMAL is closest to off).
+    const thinkingLevel =
+      isGemini3 && request.thinkingLevel === 'none'
+        ? getThinkingLevelsForModel(request.model)?.[0]
+        : request.thinkingLevel
+    if (thinkingLevel && thinkingLevel !== 'none') {
       const thinkingConfig: ThinkingConfig = { includeThoughts: request.agentEvents === true }
-      if (isGemini3OrLaterModel(model)) {
-        thinkingConfig.thinkingLevel = mapToThinkingLevel(request.thinkingLevel)
+      if (isGemini3) {
+        thinkingConfig.thinkingLevel = mapToThinkingLevel(thinkingLevel)
       } else {
-        thinkingConfig.thinkingBudget = mapToThinkingBudget(model, request.thinkingLevel)
+        thinkingConfig.thinkingBudget = mapToThinkingBudget(model, thinkingLevel)
       }
       geminiConfig.thinkingConfig = thinkingConfig
     } else if (
+      !isGemini3 &&
       request.thinkingLevel === 'none' &&
-      !isGemini3OrLaterModel(model) &&
       supportsDisablingGemini25Thinking(model)
     ) {
       // Omitting thinkingConfig falls back to the API's dynamic default (ON for gemini-2.5-flash),
@@ -1085,7 +1092,7 @@ export async function executeGeminiRequest(
      */
     const hasActiveTools = Boolean(geminiConfig.tools?.length)
     const responseFormatNeedsFinalPass =
-      Boolean(request.responseFormat) && hasActiveTools && !isGemini3OrLaterModel(model)
+      Boolean(request.responseFormat) && hasActiveTools && !isGemini3
     const liveToolLoopSupported = !responseFormatNeedsFinalPass
     const shouldStream = request.stream && !hasActiveTools
 

apps/sim/providers/models.ts

@@ -89,7 +89,8 @@ export interface ModelCapabilities {
     streamed?: ThinkingStreamVisibility
     /**
      * Explicit API mode for the pickers' `none` level on models that think by
-     * default. Omitted, `none` sends no thinking config.
+     * default. Omitted, `none` sends no thinking config, except on Gemini 3+, which
+     * has no off switch and sends its lowest declared level instead.
      */
     noneMode?: 'between_tools' | 'disabled'
     /** Whether forced tool choice can be combined with configured thinking. */