Expose reasoning effort option for reasoning models on OpenRouter

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

@@ -20,18 +20,21 @@ export const rooCodeDefaults: RooCodeSettings = {
 	// 	thinking: false,
 	// },
 
+	modelTemperature: null,
+	// reasoningEffort: "high",
+
 	pinnedApiConfigs: {},
-	lastShownAnnouncementId: "mar-20-2025-3-10",
+	lastShownAnnouncementId: "apr-04-2025-boomerang",
 
 	autoApprovalEnabled: true,
 	alwaysAllowReadOnly: true,
 	alwaysAllowReadOnlyOutsideWorkspace: false,
 	alwaysAllowWrite: true,
 	alwaysAllowWriteOutsideWorkspace: false,
-	writeDelayMs: 200,
+	writeDelayMs: 1000,
 	alwaysAllowBrowser: true,
 	alwaysApproveResubmit: true,
-	requestDelaySeconds: 5,
+	requestDelaySeconds: 10,
 	alwaysAllowMcp: true,
 	alwaysAllowModeSwitch: true,
 	alwaysAllowSubtasks: true,
@@ -40,8 +43,8 @@ export const rooCodeDefaults: RooCodeSettings = {
 
 	browserToolEnabled: false,
 	browserViewportSize: "900x600",
-	screenshotQuality: 38,
-	remoteBrowserEnabled: true,
+	screenshotQuality: 75,
+	remoteBrowserEnabled: false,
 
 	enableCheckpoints: false,
 	checkpointStorage: "task",

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

@@ -96,7 +96,7 @@ export type TelemetrySetting = z.infer<typeof telemetrySettingsSchema>
  */
 
 export const modelInfoSchema = z.object({
-	maxTokens: z.number().optional(),
+	maxTokens: z.number().nullish(),
 	contextWindow: z.number(),
 	supportsImages: z.boolean().optional(),
 	supportsComputerUse: z.boolean().optional(),
@@ -373,11 +373,14 @@ export const providerSettingsSchema = z.object({
 	requestyApiKey: z.string().optional(),
 	requestyModelId: z.string().optional(),
 	requestyModelInfo: modelInfoSchema.optional(),
-	// Generic
+	// Claude 3.7 Sonnet Thinking
 	modelMaxTokens: z.number().optional(), // Currently only used by Anthropic hybrid thinking models.
 	modelMaxThinkingTokens: z.number().optional(), // Currently only used by Anthropic hybrid thinking models.
-	modelTemperature: z.number().nullish(),
+	// Generic
 	includeMaxTokens: z.boolean().optional(),
+	modelTemperature: z.number().nullish(),
+	reasoningEffort: z.enum(["low", "medium", "high"]).optional(),
+	rateLimitSeconds: z.number().optional(),
 	// Fake AI
 	fakeAi: z.unknown().optional(),
 })
@@ -457,11 +460,13 @@ const providerSettingsRecord: ProviderSettingsRecord = {
 	requestyModelId: undefined,
 	requestyModelInfo: undefined,
 	// Claude 3.7 Sonnet Thinking
-	modelTemperature: undefined,
 	modelMaxTokens: undefined,
 	modelMaxThinkingTokens: undefined,
 	// Generic
 	includeMaxTokens: undefined,
+	modelTemperature: undefined,
+	reasoningEffort: undefined,
+	rateLimitSeconds: undefined,
 	// Fake AI
 	fakeAi: undefined,
 }

src/api/index.ts

@@ -88,21 +88,25 @@ export function getModelParams({
 	model,
 	defaultMaxTokens,
 	defaultTemperature = 0,
+	defaultReasoningEffort,
 }: {
 	options: ApiHandlerOptions
 	model: ModelInfo
 	defaultMaxTokens?: number
 	defaultTemperature?: number
+	defaultReasoningEffort?: "low" | "medium" | "high"
 }) {
 	const {
 		modelMaxTokens: customMaxTokens,
 		modelMaxThinkingTokens: customMaxThinkingTokens,
 		modelTemperature: customTemperature,
+		reasoningEffort: customReasoningEffort,
 	} = options
 
 	let maxTokens = model.maxTokens ?? defaultMaxTokens
 	let thinking: BetaThinkingConfigParam | undefined = undefined
 	let temperature = customTemperature ?? defaultTemperature
+	const reasoningEffort = customReasoningEffort ?? defaultReasoningEffort
 
 	if (model.thinking) {
 		// Only honor `customMaxTokens` for thinking models.
@@ -118,5 +122,5 @@ export function getModelParams({
 		temperature = 1.0
 	}
 
-	return { maxTokens, thinking, temperature }
+	return { maxTokens, thinking, temperature, reasoningEffort }
 }

src/api/providers/openai.ts

@@ -82,6 +82,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 		const urlHost = this._getUrlHost(modelUrl)
 		const deepseekReasoner = modelId.includes("deepseek-reasoner") || enabledR1Format
 		const ark = modelUrl.includes(".volces.com")
+
 		if (modelId.startsWith("o3-mini")) {
 			yield* this.handleO3FamilyMessage(modelId, systemPrompt, messages)
 			return
@@ -94,6 +95,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 			}
 
 			let convertedMessages
+
 			if (deepseekReasoner) {
 				convertedMessages = convertToR1Format([{ role: "user", content: systemPrompt }, ...messages])
 			} else if (ark || enabledLegacyFormat) {
@@ -112,16 +114,20 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 						],
 					}
 				}
+
 				convertedMessages = [systemMessage, ...convertToOpenAiMessages(messages)]
+
 				if (modelInfo.supportsPromptCache) {
 					// Note: the following logic is copied from openrouter:
 					// Add cache_control to the last two user messages
 					// (note: this works because we only ever add one user message at a time, but if we added multiple we'd need to mark the user message before the last assistant message)
 					const lastTwoUserMessages = convertedMessages.filter((msg) => msg.role === "user").slice(-2)
+
 					lastTwoUserMessages.forEach((msg) => {
 						if (typeof msg.content === "string") {
 							msg.content = [{ type: "text", text: msg.content }]
 						}
+
 						if (Array.isArray(msg.content)) {
 							// NOTE: this is fine since env details will always be added at the end. but if it weren't there, and the user added a image_url type message, it would pop a text part before it and then move it after to the end.
 							let lastTextPart = msg.content.filter((part) => part.type === "text").pop()
@@ -130,6 +136,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 								lastTextPart = { type: "text", text: "..." }
 								msg.content.push(lastTextPart)
 							}
+
 							// @ts-ignore-next-line
 							lastTextPart["cache_control"] = { type: "ephemeral" }
 						}
@@ -145,7 +152,9 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 				messages: convertedMessages,
 				stream: true as const,
 				...(isGrokXAI ? {} : { stream_options: { include_usage: true } }),
+				reasoning_effort: modelId === "grok-3-mini-beta" ? "high" : this.getModel().info.reasoningEffort,
 			}
+
 			if (this.options.includeMaxTokens) {
 				requestOptions.max_tokens = modelInfo.maxTokens
 			}
@@ -185,6 +194,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 					lastUsage = chunk.usage
 				}
 			}
+
 			for (const chunk of matcher.final()) {
 				yield chunk
 			}
@@ -217,6 +227,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 				type: "text",
 				text: response.choices[0]?.message.content || "",
 			}
+
 			yield this.processUsageMetrics(response.usage, modelInfo)
 		}
 	}
@@ -241,6 +252,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 	async completePrompt(prompt: string): Promise<string> {
 		try {
 			const isAzureAiInference = this._isAzureAiInference(this.options.openAiBaseUrl)
+
 			const requestOptions: OpenAI.Chat.Completions.ChatCompletionCreateParamsNonStreaming = {
 				model: this.getModel().id,
 				messages: [{ role: "user", content: prompt }],
@@ -250,11 +262,13 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 				requestOptions,
 				isAzureAiInference ? { path: AZURE_AI_INFERENCE_PATH } : {},
 			)
+
 			return response.choices[0]?.message.content || ""
 		} catch (error) {
 			if (error instanceof Error) {
 				throw new Error(`OpenAI completion error: ${error.message}`)
 			}
+
 			throw error
 		}
 	}
@@ -281,7 +295,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 					],
 					stream: true,
 					...(isGrokXAI ? {} : { stream_options: { include_usage: true } }),
-					reasoning_effort: this.getModel().info.reasoningEffort,
+					reasoning_effort: modelId === "grok-3-mini-beta" ? "high" : this.getModel().info.reasoningEffort,
 				},
 				methodIsAzureAiInference ? { path: AZURE_AI_INFERENCE_PATH } : {},
 			)
@@ -333,6 +347,7 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 			}
 		}
 	}
+
 	private _getUrlHost(baseUrl?: string): string {
 		try {
 			return new URL(baseUrl ?? "").host

src/api/providers/openrouter.ts

@@ -1,8 +1,7 @@
 import { Anthropic } from "@anthropic-ai/sdk"
 import { BetaThinkingConfigParam } from "@anthropic-ai/sdk/resources/beta"
-import axios, { AxiosRequestConfig } from "axios"
+import axios from "axios"
 import OpenAI from "openai"
-import delay from "delay"
 
 import { ApiHandlerOptions, ModelInfo, openRouterDefaultModelId, openRouterDefaultModelInfo } from "../../shared/api"
 import { parseApiPrice } from "../../utils/cost"
@@ -22,6 +21,12 @@ type OpenRouterChatCompletionParams = OpenAI.Chat.ChatCompletionCreateParams & {
 	transforms?: string[]
 	include_reasoning?: boolean
 	thinking?: BetaThinkingConfigParam
+	// https://openrouter.ai/docs/use-cases/reasoning-tokens
+	reasoning?: {
+		effort?: "high" | "medium" | "low"
+		max_tokens?: number
+		exclude?: boolean
+	}
 }
 
 export class OpenRouterHandler extends BaseProvider implements SingleCompletionHandler {
@@ -42,7 +47,7 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 		systemPrompt: string,
 		messages: Anthropic.Messages.MessageParam[],
 	): AsyncGenerator<ApiStreamChunk> {
-		let { id: modelId, maxTokens, thinking, temperature, topP } = this.getModel()
+		let { id: modelId, maxTokens, thinking, temperature, topP, reasoningEffort } = this.getModel()
 
 		// Convert Anthropic messages to OpenAI format.
 		let openAiMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
@@ -70,13 +75,16 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 						},
 					],
 				}
+
 				// Add cache_control to the last two user messages
 				// (note: this works because we only ever add one user message at a time, but if we added multiple we'd need to mark the user message before the last assistant message)
 				const lastTwoUserMessages = openAiMessages.filter((msg) => msg.role === "user").slice(-2)
+
 				lastTwoUserMessages.forEach((msg) => {
 					if (typeof msg.content === "string") {
 						msg.content = [{ type: "text", text: msg.content }]
 					}
+
 					if (Array.isArray(msg.content)) {
 						// NOTE: this is fine since env details will always be added at the end. but if it weren't there, and the user added a image_url type message, it would pop a text part before it and then move it after to the end.
 						let lastTextPart = msg.content.filter((part) => part.type === "text").pop()
@@ -113,6 +121,7 @@ export class OpenRouterHandler extends BaseProvider implements SingleCompletionH
 				}),
 			// This way, the transforms field will only be included in the parameters when openRouterUseMiddleOutTransform is true.
 			...((this.options.openRouterUseMiddleOutTransform ?? true) && { transforms: ["middle-out"] }),
+			...(reasoningEffort && { reasoning: { effort: reasoningEffort } }),
 		}
 
 		const stream = await this.client.chat.completions.create(completionParams)

src/exports/roo-code.d.ts

@@ -175,10 +175,11 @@ type ProviderSettings = {
 				cachableFields?: string[] | undefined
 		  } | null)
 		| undefined
-	modelTemperature?: (number | null) | undefined
 	modelMaxTokens?: number | undefined
 	modelMaxThinkingTokens?: number | undefined
 	includeMaxTokens?: boolean | undefined
+	modelTemperature?: (number | null) | undefined
+	reasoningEffort?: ("low" | "medium" | "high") | undefined
 	rateLimitSeconds?: number | undefined
 	fakeAi?: unknown | undefined
 }

src/exports/types.ts

@@ -176,10 +176,11 @@ type ProviderSettings = {
 				cachableFields?: string[] | undefined
 		  } | null)
 		| undefined
-	modelTemperature?: (number | null) | undefined
 	modelMaxTokens?: number | undefined
 	modelMaxThinkingTokens?: number | undefined
 	includeMaxTokens?: boolean | undefined
+	modelTemperature?: (number | null) | undefined
+	reasoningEffort?: ("low" | "medium" | "high") | undefined
 	rateLimitSeconds?: number | undefined
 	fakeAi?: unknown | undefined
 }

src/schemas/index.ts

@@ -95,6 +95,16 @@ export const telemetrySettingsSchema = z.enum(telemetrySettings)
 
 export type TelemetrySetting = z.infer<typeof telemetrySettingsSchema>
 
+/**
+ * ReasoningEffort
+ */
+
+export const reasoningEfforts = ["low", "medium", "high"] as const
+
+export const reasoningEffortsSchema = z.enum(reasoningEfforts)
+
+export type ReasoningEffort = z.infer<typeof reasoningEffortsSchema>
+
 /**
  * ModelInfo
  */
@@ -110,7 +120,7 @@ export const modelInfoSchema = z.object({
 	cacheWritesPrice: z.number().optional(),
 	cacheReadsPrice: z.number().optional(),
 	description: z.string().optional(),
-	reasoningEffort: z.enum(["low", "medium", "high"]).optional(),
+	reasoningEffort: reasoningEffortsSchema.optional(),
 	thinking: z.boolean().optional(),
 	minTokensPerCachePoint: z.number().optional(),
 	maxCachePoints: z.number().optional(),
@@ -383,11 +393,12 @@ export const providerSettingsSchema = z.object({
 	requestyModelId: z.string().optional(),
 	requestyModelInfo: modelInfoSchema.nullish(),
 	// Claude 3.7 Sonnet Thinking
-	modelTemperature: z.number().nullish(),
 	modelMaxTokens: z.number().optional(),
 	modelMaxThinkingTokens: z.number().optional(),
 	// Generic
 	includeMaxTokens: z.boolean().optional(),
+	modelTemperature: z.number().nullish(),
+	reasoningEffort: reasoningEffortsSchema.optional(),
 	rateLimitSeconds: z.number().optional(),
 	// Fake AI
 	fakeAi: z.unknown().optional(),
@@ -470,11 +481,12 @@ const providerSettingsRecord: ProviderSettingsRecord = {
 	requestyModelId: undefined,
 	requestyModelInfo: undefined,
 	// Claude 3.7 Sonnet Thinking
-	modelTemperature: undefined,
 	modelMaxTokens: undefined,
 	modelMaxThinkingTokens: undefined,
 	// Generic
 	includeMaxTokens: undefined,
+	modelTemperature: undefined,
+	reasoningEffort: undefined,
 	rateLimitSeconds: undefined,
 	// Fake AI
 	fakeAi: undefined,