feat: add manual thinking budget control for supported providers/models

evals/packages/types/src/roo-code.ts

@@ -340,7 +340,7 @@ const genericProviderSettingsSchema = z.object({
 	rateLimitSeconds: z.number().optional(),
 
 	// Model reasoning.
-	enableReasoningEffort: z.boolean().optional(),
+	setReasoningEffort: z.boolean().optional(),
 	reasoningEffort: reasoningEffortsSchema.optional(),
 	modelMaxTokens: z.number().optional(),
 	modelMaxThinkingTokens: z.number().optional(),
@@ -395,7 +395,7 @@ const openAiSchema = z.object({
 	openAiUseAzure: z.boolean().optional(),
 	azureApiVersion: z.string().optional(),
 	openAiStreamingEnabled: z.boolean().optional(),
-	enableReasoningEffort: z.boolean().optional(),
+	setReasoningEffort: z.boolean().optional(),
 	openAiHostHeader: z.string().optional(), // Keep temporarily for backward compatibility during migration.
 	openAiHeaders: z.record(z.string(), z.string()).optional(),
 })
@@ -663,7 +663,7 @@ const providerSettingsRecord: ProviderSettingsRecord = {
 	openAiUseAzure: undefined,
 	azureApiVersion: undefined,
 	openAiStreamingEnabled: undefined,
-	enableReasoningEffort: undefined,
+	setReasoningEffort: undefined,
 	openAiHostHeader: undefined, // Keep temporarily for backward compatibility during migration
 	openAiHeaders: undefined,
 	// Ollama

packages/types/src/provider-settings.ts

@@ -60,7 +60,7 @@ const baseProviderSettingsSchema = z.object({
 	rateLimitSeconds: z.number().optional(),
 
 	// Model reasoning.
-	enableReasoningEffort: z.boolean().optional(),
+	setReasoningEffort: z.boolean().optional(),
 	reasoningEffort: reasoningEffortsSchema.optional(),
 	modelMaxTokens: z.number().optional(),
 	modelMaxThinkingTokens: z.number().optional(),
@@ -331,7 +331,7 @@ export const PROVIDER_SETTINGS_KEYS = keysOf<ProviderSettings>()([
 	"codeIndexOpenAiKey",
 	"codeIndexQdrantApiKey",
 	// Reasoning
-	"enableReasoningEffort",
+	"setReasoningEffort",
 	"reasoningEffort",
 	"modelMaxTokens",
 	"modelMaxThinkingTokens",

src/api/providers/__tests__/openai.spec.ts

@@ -162,7 +162,7 @@ describe("OpenAiHandler", () => {
 		it("should include reasoning_effort when reasoning effort is enabled", async () => {
 			const reasoningOptions: ApiHandlerOptions = {
 				...mockOptions,
-				enableReasoningEffort: true,
+				setReasoningEffort: true,
 				openAiCustomModelInfo: {
 					contextWindow: 128_000,
 					supportsPromptCache: false,
@@ -184,7 +184,7 @@ describe("OpenAiHandler", () => {
 		it("should not include reasoning_effort when reasoning effort is disabled", async () => {
 			const noReasoningOptions: ApiHandlerOptions = {
 				...mockOptions,
-				enableReasoningEffort: false,
+				setReasoningEffort: false,
 				openAiCustomModelInfo: { contextWindow: 128_000, supportsPromptCache: false },
 			}
 			const noReasoningHandler = new OpenAiHandler(noReasoningOptions)

src/api/providers/gemini.ts

@@ -120,9 +120,10 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 				return {
 					id,
 					info,
-					thinkingConfig: this.options.modelMaxThinkingTokens
-						? { thinkingBudget: this.options.modelMaxThinkingTokens }
-						: undefined,
+					thinkingConfig:
+						this.options.setReasoningEffort && this.options.modelMaxThinkingTokens !== undefined
+							? { thinkingBudget: this.options.modelMaxThinkingTokens }
+							: undefined,
 					maxOutputTokens: this.options.modelMaxTokens ?? info.maxTokens ?? undefined,
 				}
 			}

src/api/providers/vertex.ts

@@ -23,9 +23,10 @@ export class VertexHandler extends GeminiHandler implements SingleCompletionHand
 				return {
 					id,
 					info,
-					thinkingConfig: this.options.modelMaxThinkingTokens
-						? { thinkingBudget: this.options.modelMaxThinkingTokens }
-						: undefined,
+					thinkingConfig:
+						this.options.setReasoningEffort && this.options.modelMaxThinkingTokens !== undefined
+							? { thinkingBudget: this.options.modelMaxThinkingTokens }
+							: undefined,
 					maxOutputTokens: this.options.modelMaxTokens ?? info.maxTokens ?? undefined,
 				}
 			}

src/api/transform/__tests__/model-params.test.ts

@@ -207,7 +207,7 @@ describe("getModelParams", () => {
 			})
 		})
 
-		it("should handle supportsReasoningBudget with enableReasoningEffort setting", () => {
+		it("should handle supportsReasoningBudget with setReasoningEffort setting", () => {
 			const model: ModelInfo = {
 				...baseModel,
 				maxTokens: 2000,
@@ -216,7 +216,7 @@ describe("getModelParams", () => {
 
 			const result = getModelParams({
 				...anthropicParams,
-				settings: { enableReasoningEffort: true },
+				settings: { setReasoningEffort: true },
 				model,
 			})
 
@@ -228,7 +228,7 @@ describe("getModelParams", () => {
 			})
 		})
 
-		it("should not use reasoning budget when supportsReasoningBudget is true but enableReasoningEffort is false", () => {
+		it("should not use reasoning budget when supportsReasoningBudget is true but setReasoningEffort is false", () => {
 			const model: ModelInfo = {
 				...baseModel,
 				maxTokens: 2000,
@@ -237,7 +237,7 @@ describe("getModelParams", () => {
 
 			const result = getModelParams({
 				...anthropicParams,
-				settings: { enableReasoningEffort: false },
+				settings: { setReasoningEffort: false },
 				model,
 			})
 
@@ -537,7 +537,7 @@ describe("getModelParams", () => {
 		it("should keep model maxTokens for hybrid models when using reasoning budget", () => {
 			const result = getModelParams({
 				...anthropicParams,
-				settings: { enableReasoningEffort: true },
+				settings: { setReasoningEffort: true },
 				model,
 			})
 
@@ -560,7 +560,7 @@ describe("getModelParams", () => {
 			// Only reasoning budget should be used (takes precedence)
 			const result = getModelParams({
 				...anthropicParams,
-				settings: { enableReasoningEffort: true },
+				settings: { setReasoningEffort: true },
 				model,
 			})
 
@@ -645,7 +645,7 @@ describe("getModelParams", () => {
 			const result = getModelParams({
 				...anthropicParams,
 				settings: {
-					enableReasoningEffort: true,
+					setReasoningEffort: true,
 					modelMaxTokens: 20000,
 					modelMaxThinkingTokens: 10000,
 					modelTemperature: 0.8,

src/api/transform/__tests__/reasoning.test.ts

@@ -47,7 +47,7 @@ describe("reasoning.ts", () => {
 			}
 
 			const settingsWithEnabled: ProviderSettings = {
-				enableReasoningEffort: true,
+				setReasoningEffort: true,
 			}
 
 			const options = {
@@ -108,7 +108,7 @@ describe("reasoning.ts", () => {
 			}
 
 			const settingsWithBoth: ProviderSettings = {
-				enableReasoningEffort: true,
+				setReasoningEffort: true,
 				reasoningEffort: "low",
 			}
 
@@ -189,14 +189,14 @@ describe("reasoning.ts", () => {
 			expect(result).toEqual({ max_tokens: 0 })
 		})
 
-		it("should not use reasoning budget when supportsReasoningBudget is true but enableReasoningEffort is false", () => {
+		it("should not use reasoning budget when supportsReasoningBudget is true but setReasoningEffort is false", () => {
 			const modelWithSupported: ModelInfo = {
 				...baseModel,
 				supportsReasoningBudget: true,
 			}
 
 			const settingsWithDisabled: ProviderSettings = {
-				enableReasoningEffort: false,
+				setReasoningEffort: false,
 			}
 
 			const options = {
@@ -252,7 +252,7 @@ describe("reasoning.ts", () => {
 			}
 
 			const settingsWithEnabled: ProviderSettings = {
-				enableReasoningEffort: true,
+				setReasoningEffort: true,
 			}
 
 			const options = {
@@ -274,14 +274,14 @@ describe("reasoning.ts", () => {
 			expect(result).toBeUndefined()
 		})
 
-		it("should return undefined when supportsReasoningBudget is true but enableReasoningEffort is false", () => {
+		it("should return undefined when supportsReasoningBudget is true but setReasoningEffort is false", () => {
 			const modelWithSupported: ModelInfo = {
 				...baseModel,
 				supportsReasoningBudget: true,
 			}
 
 			const settingsWithDisabled: ProviderSettings = {
-				enableReasoningEffort: false,
+				setReasoningEffort: false,
 			}
 
 			const options = {
@@ -513,7 +513,7 @@ describe("reasoning.ts", () => {
 			}
 
 			const settingsWithEnabled: ProviderSettings = {
-				enableReasoningEffort: true,
+				setReasoningEffort: true,
 			}
 
 			const options = {
@@ -583,7 +583,7 @@ describe("reasoning.ts", () => {
 			}
 
 			const settingsWithBoth: ProviderSettings = {
-				enableReasoningEffort: true,
+				setReasoningEffort: true,
 				reasoningEffort: "high",
 			}
 

src/api/transform/model-params.ts

@@ -1,7 +1,7 @@
 import type { ModelInfo, ProviderSettings } from "@roo-code/types"
 
 import { ANTHROPIC_DEFAULT_MAX_TOKENS } from "../providers/constants"
-import { shouldUseReasoningBudget, shouldUseReasoningEffort } from "../../shared/api"
+import { shouldSetReasoningBudget, shouldUseReasoningEffort } from "../../shared/api"
 
 import {
 	type AnthropicReasoningParams,
@@ -67,7 +67,7 @@ export function getModelParams({
 	let reasoningBudget: ModelParams["reasoningBudget"] = undefined
 	let reasoningEffort: ModelParams["reasoningEffort"] = undefined
 
-	if (shouldUseReasoningBudget({ model, settings })) {
+	if (shouldSetReasoningBudget({ model, settings })) {
 		// "Hybrid" reasoning models use the `reasoningBudget` parameter.
 		maxTokens = customMaxTokens ?? maxTokens
 

src/api/transform/reasoning.ts

@@ -3,7 +3,7 @@ import OpenAI from "openai"
 
 import type { ModelInfo, ProviderSettings } from "@roo-code/types"
 
-import { shouldUseReasoningBudget, shouldUseReasoningEffort } from "../../shared/api"
+import { shouldSetReasoningBudget, shouldUseReasoningEffort } from "../../shared/api"
 
 type ReasoningEffort = "low" | "medium" | "high"
 
@@ -30,7 +30,7 @@ export const getOpenRouterReasoning = ({
 	reasoningEffort,
 	settings,
 }: GetModelReasoningOptions): OpenRouterReasoningParams | undefined =>
-	shouldUseReasoningBudget({ model, settings })
+	shouldSetReasoningBudget({ model, settings })
 		? { max_tokens: reasoningBudget }
 		: shouldUseReasoningEffort({ model, settings })
 			? { effort: reasoningEffort }
@@ -41,7 +41,7 @@ export const getAnthropicReasoning = ({
 	reasoningBudget,
 	settings,
 }: GetModelReasoningOptions): AnthropicReasoningParams | undefined =>
-	shouldUseReasoningBudget({ model, settings }) ? { type: "enabled", budget_tokens: reasoningBudget! } : undefined
+	shouldSetReasoningBudget({ model, settings }) ? { type: "enabled", budget_tokens: reasoningBudget! } : undefined
 
 export const getOpenAiReasoning = ({
 	model,

src/core/config/ProviderSettingsManager.ts

@@ -26,6 +26,7 @@ export const providerProfilesSchema = z.object({
 			rateLimitSecondsMigrated: z.boolean().optional(),
 			diffSettingsMigrated: z.boolean().optional(),
 			openAiHeadersMigrated: z.boolean().optional(),
+			manualThinkingBudgetMigrated: z.boolean().optional(),
 		})
 		.optional(),
 })
@@ -48,6 +49,7 @@ export class ProviderSettingsManager {
 			rateLimitSecondsMigrated: true, // Mark as migrated on fresh installs
 			diffSettingsMigrated: true, // Mark as migrated on fresh installs
 			openAiHeadersMigrated: true, // Mark as migrated on fresh installs
+			manualThinkingBudgetMigrated: true, // Mark as migrated on fresh installs
 		},
 	}
 
@@ -113,6 +115,7 @@ export class ProviderSettingsManager {
 						rateLimitSecondsMigrated: false,
 						diffSettingsMigrated: false,
 						openAiHeadersMigrated: false,
+						manualThinkingBudgetMigrated: false,
 					} // Initialize with default values
 					isDirty = true
 				}
@@ -135,6 +138,12 @@ export class ProviderSettingsManager {
 					isDirty = true
 				}
 
+				if (!providerProfiles.migrations.manualThinkingBudgetMigrated) {
+					await this.migrateManualThinkingBudget(providerProfiles)
+					providerProfiles.migrations.manualThinkingBudgetMigrated = true
+					isDirty = true
+				}
+
 				if (isDirty) {
 					await this.store(providerProfiles)
 				}
@@ -228,6 +237,23 @@ export class ProviderSettingsManager {
 		}
 	}
 
+	private async migrateManualThinkingBudget(providerProfiles: ProviderProfiles) {
+		try {
+			for (const [_name, apiConfig] of Object.entries(providerProfiles.apiConfigs)) {
+				// For existing users who have modelMaxThinkingTokens set, enable manual control
+				// This maintains backward compatibility - if they were manually setting thinking tokens,
+				// they should continue to have manual control enabled
+				if (apiConfig.modelMaxThinkingTokens !== undefined && apiConfig.setReasoningEffort === undefined) {
+					apiConfig.setReasoningEffort = true
+				}
+				// For new users or existing users without thinking tokens set,
+				// default to false (automatic mode) - this is handled by the UI component's default logic
+			}
+		} catch (error) {
+			console.error(`[MigrateManualThinkingBudget] Failed to migrate manual thinking budget settings:`, error)
+		}
+	}
+
 	/**
 	 * List all available configs with metadata.
 	 */