Feat/issue 5784 custom max tokens

src/api/providers/__tests__/anthropic-vertex.spec.ts

@@ -705,7 +705,7 @@ describe("VertexHandler", () => {
 			expect(result.temperature).toBe(1.0)
 		})
 
-		it("does not honor custom maxTokens for non-thinking models", () => {
+		it("honors custom maxTokens for all models", () => {
 			const handler = new AnthropicVertexHandler({
 				apiKey: "test-api-key",
 				apiModelId: "claude-3-7-sonnet@20250219",
@@ -714,7 +714,7 @@ describe("VertexHandler", () => {
 			})
 
 			const result = handler.getModel()
-			expect(result.maxTokens).toBe(8192)
+			expect(result.maxTokens).toBe(8192) // Capped to model's actual maxTokens
 			expect(result.reasoningBudget).toBeUndefined()
 			expect(result.temperature).toBe(0)
 		})

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

@@ -251,7 +251,7 @@ describe("AnthropicHandler", () => {
 			expect(result.temperature).toBe(1.0)
 		})
 
-		it("does not honor custom maxTokens for non-thinking models", () => {
+		it("honors custom maxTokens for all models", () => {
 			const handler = new AnthropicHandler({
 				apiKey: "test-api-key",
 				apiModelId: "claude-3-7-sonnet-20250219",
@@ -260,7 +260,7 @@ describe("AnthropicHandler", () => {
 			})
 
 			const result = handler.getModel()
-			expect(result.maxTokens).toBe(8192)
+			expect(result.maxTokens).toBe(8192) // Capped to model's actual maxTokens
 			expect(result.reasoningBudget).toBeUndefined()
 			expect(result.temperature).toBe(0)
 		})

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

@@ -263,14 +263,14 @@ describe("OpenAiHandler", () => {
 			expect(callArgs.max_completion_tokens).toBeUndefined()
 		})
 
-		it("should use user-configured modelMaxTokens instead of model default maxTokens", async () => {
+		it("should use user-configured modelMaxTokens but cap it to model's max capability", async () => {
 			const optionsWithUserMaxTokens: ApiHandlerOptions = {
 				...mockOptions,
 				includeMaxTokens: true,
-				modelMaxTokens: 32000, // User-configured value
+				modelMaxTokens: 32000, // User tries to set higher than model supports
 				openAiCustomModelInfo: {
 					contextWindow: 128_000,
-					maxTokens: 4096, // Model's default value (should not be used)
+					maxTokens: 4096, // Model's actual max capability
 					supportsPromptCache: false,
 				},
 			}
@@ -279,10 +279,32 @@ describe("OpenAiHandler", () => {
 			// Consume the stream to trigger the API call
 			for await (const _chunk of stream) {
 			}
-			// Assert the mockCreate was called with user-configured modelMaxTokens (32000), not model default maxTokens (4096)
+			// Assert the mockCreate was called with the model's max capability (4096), not the user's request (32000)
 			expect(mockCreate).toHaveBeenCalled()
 			const callArgs = mockCreate.mock.calls[0][0]
-			expect(callArgs.max_completion_tokens).toBe(32000)
+			expect(callArgs.max_completion_tokens).toBe(4096)
+		})
+
+		it("should use user-configured modelMaxTokens when it's less than model's max", async () => {
+			const optionsWithLowerUserMaxTokens: ApiHandlerOptions = {
+				...mockOptions,
+				includeMaxTokens: true,
+				modelMaxTokens: 2000, // User sets lower than model's max
+				openAiCustomModelInfo: {
+					contextWindow: 128_000,
+					maxTokens: 4096, // Model's max capability
+					supportsPromptCache: false,
+				},
+			}
+			const handlerWithLowerMaxTokens = new OpenAiHandler(optionsWithLowerUserMaxTokens)
+			const stream = handlerWithLowerMaxTokens.createMessage(systemPrompt, messages)
+			// Consume the stream to trigger the API call
+			for await (const _chunk of stream) {
+			}
+			// Assert the mockCreate was called with user's setting (2000)
+			expect(mockCreate).toHaveBeenCalled()
+			const callArgs = mockCreate.mock.calls[0][0]
+			expect(callArgs.max_completion_tokens).toBe(2000)
 		})
 
 		it("should fallback to model default maxTokens when user modelMaxTokens is not set", async () => {

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

@@ -89,7 +89,7 @@ describe("OpenRouterHandler", () => {
 			expect(result.info.supportsPromptCache).toBe(true)
 		})
 
-		it("honors custom maxTokens for thinking models", async () => {
+		it("honors custom maxTokens for all models", async () => {
 			const handler = new OpenRouterHandler({
 				openRouterApiKey: "test-key",
 				openRouterModelId: "anthropic/claude-3.7-sonnet:thinking",
@@ -98,20 +98,20 @@ describe("OpenRouterHandler", () => {
 			})
 
 			const result = await handler.fetchModel()
-			expect(result.maxTokens).toBe(128000) // Use actual implementation value
+			expect(result.maxTokens).toBe(32_768) // Not capped since model supports 128000
 			expect(result.reasoningBudget).toBeUndefined() // Use actual implementation value
 			expect(result.temperature).toBe(0) // Use actual implementation value
 		})
 
-		it("does not honor custom maxTokens for non-thinking models", async () => {
+		it("honors custom maxTokens for non-thinking models", async () => {
 			const handler = new OpenRouterHandler({
 				...mockOptions,
 				modelMaxTokens: 32_768,
 				modelMaxThinkingTokens: 16_384,
 			})
 
 			const result = await handler.fetchModel()
-			expect(result.maxTokens).toBe(8192)
+			expect(result.maxTokens).toBe(8192) // Capped to model's actual maxTokens
 			expect(result.reasoningBudget).toBeUndefined()
 			expect(result.temperature).toBe(0)
 		})

src/api/providers/chutes.ts

@@ -3,6 +3,7 @@ import { Anthropic } from "@anthropic-ai/sdk"
 import OpenAI from "openai"
 
 import type { ApiHandlerOptions } from "../../shared/api"
+import { getModelMaxOutputTokens } from "../../shared/api"
 import { XmlMatcher } from "../../utils/xml-matcher"
 import { convertToR1Format } from "../transform/r1-format"
 import { convertToOpenAiMessages } from "../transform/openai-format"
@@ -27,15 +28,17 @@ export class ChutesHandler extends BaseOpenAiCompatibleProvider<ChutesModelId> {
 		systemPrompt: string,
 		messages: Anthropic.Messages.MessageParam[],
 	): OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming {
-		const {
-			id: model,
-			info: { maxTokens: max_tokens },
-		} = this.getModel()
+		const model = this.getModel()
+		const max_tokens = getModelMaxOutputTokens({
+			modelId: model.id,
+			model: model.info,
+			settings: this.options as any,
+		})
 
-		const temperature = this.options.modelTemperature ?? this.getModel().info.temperature
+		const temperature = this.options.modelTemperature ?? model.info.temperature
 
 		return {
-			model,
+			model: model.id,
 			max_tokens,
 			temperature,
 			messages: [{ role: "system", content: systemPrompt }, ...convertToOpenAiMessages(messages)],

src/api/providers/gemini.ts

@@ -71,7 +71,7 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 			systemInstruction,
 			httpOptions: this.options.googleGeminiBaseUrl ? { baseUrl: this.options.googleGeminiBaseUrl } : undefined,
 			thinkingConfig,
-			maxOutputTokens: this.options.modelMaxTokens ?? maxTokens ?? undefined,
+			maxOutputTokens: maxTokens ?? undefined,
 			temperature: this.options.modelTemperature ?? 0,
 		}
 

src/api/providers/glama.ts

@@ -5,7 +5,7 @@ import OpenAI from "openai"
 import { glamaDefaultModelId, glamaDefaultModelInfo, GLAMA_DEFAULT_TEMPERATURE } from "@roo-code/types"
 
 import { Package } from "../../shared/package"
-import { ApiHandlerOptions } from "../../shared/api"
+import { ApiHandlerOptions, getModelMaxOutputTokens } from "../../shared/api"
 
 import { ApiStream } from "../transform/stream"
 import { convertToOpenAiMessages } from "../transform/openai-format"
@@ -49,12 +49,14 @@ export class GlamaHandler extends RouterProvider implements SingleCompletionHand
 			addCacheBreakpoints(systemPrompt, openAiMessages)
 		}
 
-		// Required by Anthropic; other providers default to max tokens allowed.
-		let maxTokens: number | undefined
-
-		if (modelId.startsWith("anthropic/")) {
-			maxTokens = info.maxTokens ?? undefined
-		}
+		// Use getModelMaxOutputTokens to respect user's custom max tokens setting
+		const maxTokens = modelId.startsWith("anthropic/")
+			? getModelMaxOutputTokens({
+					modelId,
+					model: info,
+					settings: this.options as any,
+				})
+			: undefined
 
 		const requestOptions: OpenAI.Chat.ChatCompletionCreateParams = {
 			model: modelId,
@@ -130,7 +132,11 @@ export class GlamaHandler extends RouterProvider implements SingleCompletionHand
 			}
 
 			if (modelId.startsWith("anthropic/")) {
-				requestOptions.max_tokens = info.maxTokens
+				requestOptions.max_tokens = getModelMaxOutputTokens({
+					modelId,
+					model: info,
+					settings: this.options as any,
+				})
 			}
 
 			const response = await this.client.chat.completions.create(requestOptions)

src/api/providers/lite-llm.ts

@@ -5,7 +5,7 @@ import { litellmDefaultModelId, litellmDefaultModelInfo } from "@roo-code/types"
 
 import { calculateApiCostOpenAI } from "../../shared/cost"
 
-import { ApiHandlerOptions } from "../../shared/api"
+import { ApiHandlerOptions, getModelMaxOutputTokens } from "../../shared/api"
 
 import { ApiStream, ApiStreamUsageChunk } from "../transform/stream"
 import { convertToOpenAiMessages } from "../transform/openai-format"
@@ -44,8 +44,12 @@ export class LiteLLMHandler extends RouterProvider implements SingleCompletionHa
 			...convertToOpenAiMessages(messages),
 		]
 
-		// Required by some providers; others default to max tokens allowed
-		let maxTokens: number | undefined = info.maxTokens ?? undefined
+		// Use getModelMaxOutputTokens to respect user's custom max tokens setting
+		const maxTokens = getModelMaxOutputTokens({
+			modelId,
+			model: info,
+			settings: this.options as any,
+		})
 
 		const requestOptions: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {
 			model: modelId,
@@ -119,7 +123,11 @@ export class LiteLLMHandler extends RouterProvider implements SingleCompletionHa
 				requestOptions.temperature = this.options.modelTemperature ?? 0
 			}
 
-			requestOptions.max_tokens = info.maxTokens
+			requestOptions.max_tokens = getModelMaxOutputTokens({
+				modelId,
+				model: info,
+				settings: this.options as any,
+			})
 
 			const response = await this.client.chat.completions.create(requestOptions)
 			return response.choices[0]?.message.content || ""

src/api/providers/mistral.ts

@@ -3,7 +3,7 @@ import { Mistral } from "@mistralai/mistralai"
 
 import { type MistralModelId, mistralDefaultModelId, mistralModels, MISTRAL_DEFAULT_TEMPERATURE } from "@roo-code/types"
 
-import { ApiHandlerOptions } from "../../shared/api"
+import { ApiHandlerOptions, getModelMaxOutputTokens } from "../../shared/api"
 
 import { convertToMistralMessages } from "../transform/mistral-format"
 import { ApiStream } from "../transform/stream"
@@ -78,7 +78,13 @@ export class MistralHandler extends BaseProvider implements SingleCompletionHand
 		const info = mistralModels[id as MistralModelId] ?? mistralModels[mistralDefaultModelId]
 
 		// @TODO: Move this to the `getModelParams` function.
-		const maxTokens = this.options.includeMaxTokens ? info.maxTokens : undefined
+		const maxTokens = this.options.includeMaxTokens
+			? getModelMaxOutputTokens({
+					modelId: id,
+					model: info,
+					settings: { ...this.options, apiProvider: "mistral" } as any,
+				})
+			: undefined
 		const temperature = this.options.modelTemperature ?? MISTRAL_DEFAULT_TEMPERATURE
 
 		return { id, info, maxTokens, temperature }

src/api/providers/openai.ts

@@ -10,7 +10,7 @@ import {
 	OPENAI_AZURE_AI_INFERENCE_PATH,
 } from "@roo-code/types"
 
-import type { ApiHandlerOptions } from "../../shared/api"
+import { ApiHandlerOptions, getModelMaxOutputTokens } from "../../shared/api"
 
 import { XmlMatcher } from "../../utils/xml-matcher"
 
@@ -403,9 +403,14 @@ export class OpenAiHandler extends BaseProvider implements SingleCompletionHandl
 	): void {
 		// Only add max_completion_tokens if includeMaxTokens is true
 		if (this.options.includeMaxTokens === true) {
-			// Use user-configured modelMaxTokens if available, otherwise fall back to model's default maxTokens
+			// Use getModelMaxOutputTokens to properly handle user settings and model limits
 			// Using max_completion_tokens as max_tokens is deprecated
-			requestOptions.max_completion_tokens = this.options.modelMaxTokens || modelInfo.maxTokens
+			const modelId = this.options.openAiModelId ?? ""
+			requestOptions.max_completion_tokens = getModelMaxOutputTokens({
+				modelId,
+				model: modelInfo,
+				settings: { ...this.options, apiProvider: "openai" } as any,
+			})
 		}
 	}
 }

src/api/providers/unbound.ts

@@ -4,6 +4,7 @@ import OpenAI from "openai"
 import { unboundDefaultModelId, unboundDefaultModelInfo } from "@roo-code/types"
 
 import type { ApiHandlerOptions } from "../../shared/api"
+import { getModelMaxOutputTokens } from "../../shared/api"
 
 import { ApiStream, ApiStreamUsageChunk } from "../transform/stream"
 import { convertToOpenAiMessages } from "../transform/openai-format"
@@ -76,12 +77,14 @@ export class UnboundHandler extends RouterProvider implements SingleCompletionHa
 			addVertexCacheBreakpoints(messages)
 		}
 
-		// Required by Anthropic; other providers default to max tokens allowed.
-		let maxTokens: number | undefined
-
-		if (modelId.startsWith("anthropic/")) {
-			maxTokens = info.maxTokens ?? undefined
-		}
+		// Use getModelMaxOutputTokens to respect user's custom max tokens setting
+		const maxTokens = modelId.startsWith("anthropic/")
+			? getModelMaxOutputTokens({
+					modelId,
+					model: info,
+					settings: this.options as any,
+				})
+			: undefined
 
 		const requestOptions: UnboundChatCompletionCreateParamsStreaming = {
 			model: modelId.split("/")[1],
@@ -149,7 +152,11 @@ export class UnboundHandler extends RouterProvider implements SingleCompletionHa
 			}
 
 			if (modelId.startsWith("anthropic/")) {
-				requestOptions.max_tokens = info.maxTokens
+				requestOptions.max_tokens = getModelMaxOutputTokens({
+					modelId,
+					model: info,
+					settings: this.options as any,
+				})
 			}
 
 			const response = await this.client.chat.completions.create(requestOptions, { headers: DEFAULT_HEADERS })

src/api/providers/xai.ts

@@ -3,7 +3,7 @@ import OpenAI from "openai"
 
 import { type XAIModelId, xaiDefaultModelId, xaiModels } from "@roo-code/types"
 
-import type { ApiHandlerOptions } from "../../shared/api"
+import { ApiHandlerOptions, getModelMaxOutputTokens } from "../../shared/api"
 
 import { ApiStream } from "../transform/stream"
 import { convertToOpenAiMessages } from "../transform/openai-format"
@@ -50,7 +50,11 @@ export class XAIHandler extends BaseProvider implements SingleCompletionHandler
 		// Use the OpenAI-compatible API.
 		const stream = await this.client.chat.completions.create({
 			model: modelId,
-			max_tokens: modelInfo.maxTokens,
+			max_tokens: getModelMaxOutputTokens({
+				modelId,
+				model: modelInfo,
+				settings: { ...this.options, apiProvider: "xai" } as any,
+			}),
 			temperature: this.options.modelTemperature ?? XAI_DEFAULT_TEMPERATURE,
 			messages: [{ role: "system", content: systemPrompt }, ...convertToOpenAiMessages(messages)],
 			stream: true,
@@ -78,12 +82,15 @@ export class XAIHandler extends BaseProvider implements SingleCompletionHandler
 			if (chunk.usage) {
 				// Extract detailed token information if available
 				// First check for prompt_tokens_details structure (real API response)
-				const promptDetails = "prompt_tokens_details" in chunk.usage ? chunk.usage.prompt_tokens_details : null;
-				const cachedTokens = promptDetails && "cached_tokens" in promptDetails ? promptDetails.cached_tokens : 0;
+				const promptDetails = "prompt_tokens_details" in chunk.usage ? chunk.usage.prompt_tokens_details : null
+				const cachedTokens = promptDetails && "cached_tokens" in promptDetails ? promptDetails.cached_tokens : 0
 
 				// Fall back to direct fields in usage (used in test mocks)
-				const readTokens = cachedTokens || ("cache_read_input_tokens" in chunk.usage ? (chunk.usage as any).cache_read_input_tokens : 0);
-				const writeTokens = "cache_creation_input_tokens" in chunk.usage ? (chunk.usage as any).cache_creation_input_tokens : 0;
+				const readTokens =
+					cachedTokens ||
+					("cache_read_input_tokens" in chunk.usage ? (chunk.usage as any).cache_read_input_tokens : 0)
+				const writeTokens =
+					"cache_creation_input_tokens" in chunk.usage ? (chunk.usage as any).cache_creation_input_tokens : 0
 
 				yield {
 					type: "usage",

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

@@ -665,7 +665,7 @@ describe("getModelParams", () => {
 				model,
 			})
 
-			expect(result.maxTokens).toBe(20000)
+			expect(result.maxTokens).toBe(16000) // Capped to model's actual maxTokens
 			expect(result.reasoningBudget).toBe(10000)
 			expect(result.temperature).toBe(1.0) // Overridden for reasoning budget models
 			expect(result.reasoningEffort).toBeUndefined() // Budget takes precedence