Re-work support for reasoning models, including "hybrid" reasoning models (#3870)

Author: cte

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

@@ -335,12 +335,14 @@ export type ProviderSettingsEntry = z.infer<typeof providerSettingsEntrySchema>
 
 const genericProviderSettingsSchema = z.object({
 	includeMaxTokens: z.boolean().optional(),
-	reasoningEffort: reasoningEffortsSchema.optional(),
 	diffEnabled: z.boolean().optional(),
 	fuzzyMatchThreshold: z.number().optional(),
 	modelTemperature: z.number().nullish(),
 	rateLimitSeconds: z.number().optional(),
-	// Claude 3.7 Sonnet Thinking
+
+	// Model reasoning.
+	enableReasoningEffort: z.boolean().optional(),
+	reasoningEffort: reasoningEffortsSchema.optional(),
 	modelMaxTokens: z.number().optional(),
 	modelMaxThinkingTokens: z.number().optional(),
 })

package.json

@@ -14,6 +14,7 @@
 		"clean": "turbo clean --log-order grouped --output-logs new-only && rimraf dist out bin .vite-port .turbo",
 		"build": "pnpm --filter roo-cline vsix",
 		"build:nightly": "pnpm --filter @roo-code/vscode-nightly vsix",
+		"generate-types": "pnpm --filter roo-cline generate-types",
 		"changeset:version": "cp CHANGELOG.md src/CHANGELOG.md && changeset version && cp -vf src/CHANGELOG.md .",
 		"knip": "pnpm --filter @roo-code/build build && knip --include files",
 		"update-contributors": "node scripts/update-contributors.js"

src/api/__tests__/index.test.ts

@@ -1,8 +1,6 @@
 import { Anthropic } from "@anthropic-ai/sdk"
-import { BetaThinkingConfigParam } from "@anthropic-ai/sdk/resources/beta/messages/index.mjs"
 
-import { ProviderSettings, ModelInfo, ApiHandlerOptions } from "../shared/api"
-import { ANTHROPIC_DEFAULT_MAX_TOKENS } from "./providers/constants"
+import { ProviderSettings, ModelInfo } from "../shared/api"
 import { GlamaHandler } from "./providers/glama"
 import { AnthropicHandler } from "./providers/anthropic"
 import { AwsBedrockHandler } from "./providers/bedrock"
@@ -101,45 +99,3 @@ export function buildApiHandler(configuration: ProviderSettings): ApiHandler {
 			return new AnthropicHandler(options)
 	}
 }
-
-export function getModelParams({
-	options,
-	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.
-		maxTokens = customMaxTokens ?? maxTokens
-
-		// Clamp the thinking budget to be at most 80% of max tokens and at
-		// least 1024 tokens.
-		const maxBudgetTokens = Math.floor((maxTokens || ANTHROPIC_DEFAULT_MAX_TOKENS) * 0.8)
-		const budgetTokens = Math.max(Math.min(customMaxThinkingTokens ?? maxBudgetTokens, maxBudgetTokens), 1024)
-		thinking = { type: "enabled", budget_tokens: budgetTokens }
-
-		// Anthropic "Thinking" models require a temperature of 1.0.
-		temperature = 1.0
-	}
-
-	return { maxTokens, thinking, temperature, reasoningEffort }
-}

src/api/index.ts

@@ -701,7 +701,7 @@ describe("VertexHandler", () => {
 
 			const result = handler.getModel()
 			expect(result.maxTokens).toBe(32_768)
-			expect(result.thinking).toEqual({ type: "enabled", budget_tokens: 16_384 })
+			expect(result.reasoningBudget).toEqual(16_384)
 			expect(result.temperature).toBe(1.0)
 		})
 
@@ -715,7 +715,7 @@ describe("VertexHandler", () => {
 
 			const result = handler.getModel()
 			expect(result.maxTokens).toBe(8192)
-			expect(result.thinking).toBeUndefined()
+			expect(result.reasoningBudget).toBeUndefined()
 			expect(result.temperature).toBe(0)
 		})
 	})
@@ -732,13 +732,9 @@ describe("VertexHandler", () => {
 
 			const modelInfo = thinkingHandler.getModel()
 
-			// Verify thinking configuration
 			expect(modelInfo.id).toBe("claude-3-7-sonnet@20250219")
-			expect(modelInfo.thinking).toBeDefined()
-			const thinkingConfig = modelInfo.thinking as { type: "enabled"; budget_tokens: number }
-			expect(thinkingConfig.type).toBe("enabled")
-			expect(thinkingConfig.budget_tokens).toBe(4096)
-			expect(modelInfo.temperature).toBe(1.0) // Thinking requires temperature 1.0
+			expect(modelInfo.reasoningBudget).toBe(4096)
+			expect(modelInfo.temperature).toBe(1.0) // Thinking requires temperature 1.0.
 		})
 
 		it("should calculate thinking budget correctly", () => {
@@ -751,7 +747,7 @@ describe("VertexHandler", () => {
 				modelMaxThinkingTokens: 5000,
 			})
 
-			expect((handlerWithBudget.getModel().thinking as any).budget_tokens).toBe(5000)
+			expect(handlerWithBudget.getModel().reasoningBudget).toBe(5000)
 
 			// Test with default thinking budget (80% of max tokens)
 			const handlerWithDefaultBudget = new AnthropicVertexHandler({
@@ -761,7 +757,7 @@ describe("VertexHandler", () => {
 				modelMaxTokens: 10000,
 			})
 
-			expect((handlerWithDefaultBudget.getModel().thinking as any).budget_tokens).toBe(8000) // 80% of 10000
+			expect(handlerWithDefaultBudget.getModel().reasoningBudget).toBe(8000) // 80% of 10000
 
 			// Test with minimum thinking budget (should be at least 1024)
 			const handlerWithSmallMaxTokens = new AnthropicVertexHandler({
@@ -771,7 +767,7 @@ describe("VertexHandler", () => {
 				modelMaxTokens: 1000, // This would result in 800 tokens for thinking, but minimum is 1024
 			})
 
-			expect((handlerWithSmallMaxTokens.getModel().thinking as any).budget_tokens).toBe(1024)
+			expect(handlerWithSmallMaxTokens.getModel().reasoningBudget).toBe(1024)
 		})
 
 		it("should pass thinking configuration to API", async () => {