feat(claude-code): add configurable max output tokens setting (#5610)

Co-authored-by: Copilot <[email protected]>
Co-authored-by: Daniel <[email protected]>
Co-authored-by: Daniel Riccio <[email protected]>

Author: 3 people

packages/types/src/provider-settings.ts

@@ -80,6 +80,7 @@ const anthropicSchema = apiModelIdProviderModelSchema.extend({
 
 const claudeCodeSchema = apiModelIdProviderModelSchema.extend({
 	claudeCodePath: z.string().optional(),
+	claudeCodeMaxOutputTokens: z.number().int().min(1).max(200000).optional(),
 })
 
 const glamaSchema = baseProviderSettingsSchema.extend({

packages/types/src/providers/claude-code.ts

@@ -4,6 +4,7 @@ import { anthropicModels } from "./anthropic.js"
 // Claude Code
 export type ClaudeCodeModelId = keyof typeof claudeCodeModels
 export const claudeCodeDefaultModelId: ClaudeCodeModelId = "claude-sonnet-4-20250514"
+export const CLAUDE_CODE_DEFAULT_MAX_OUTPUT_TOKENS = 8000
 export const claudeCodeModels = {
 	"claude-sonnet-4-20250514": {
 		...anthropicModels["claude-sonnet-4-20250514"],

src/api/providers/__tests__/claude-code.spec.ts

@@ -48,6 +48,32 @@ describe("ClaudeCodeHandler", () => {
 		expect(model.id).toBe("claude-sonnet-4-20250514") // default model
 	})
 
+	test("should override maxTokens when claudeCodeMaxOutputTokens is provided", () => {
+		const options: ApiHandlerOptions = {
+			claudeCodePath: "claude",
+			apiModelId: "claude-sonnet-4-20250514",
+			claudeCodeMaxOutputTokens: 8000,
+		}
+		const handlerWithMaxTokens = new ClaudeCodeHandler(options)
+		const model = handlerWithMaxTokens.getModel()
+
+		expect(model.id).toBe("claude-sonnet-4-20250514")
+		expect(model.info.maxTokens).toBe(8000) // Should use the configured value, not the default 64000
+	})
+
+	test("should override maxTokens for default model when claudeCodeMaxOutputTokens is provided", () => {
+		const options: ApiHandlerOptions = {
+			claudeCodePath: "claude",
+			apiModelId: "invalid-model", // Will fall back to default
+			claudeCodeMaxOutputTokens: 16384,
+		}
+		const handlerWithMaxTokens = new ClaudeCodeHandler(options)
+		const model = handlerWithMaxTokens.getModel()
+
+		expect(model.id).toBe("claude-sonnet-4-20250514") // default model
+		expect(model.info.maxTokens).toBe(16384) // Should use the configured value
+	})
+
 	test("should filter messages and call runClaudeCode", async () => {
 		const systemPrompt = "You are a helpful assistant"
 		const messages = [{ role: "user" as const, content: "Hello" }]
@@ -76,6 +102,43 @@ describe("ClaudeCodeHandler", () => {
 			messages: filteredMessages,
 			path: "claude",
 			modelId: "claude-3-5-sonnet-20241022",
+			maxOutputTokens: undefined, // No maxOutputTokens configured in this test
+		})
+	})
+
+	test("should pass maxOutputTokens to runClaudeCode when configured", async () => {
+		const options: ApiHandlerOptions = {
+			claudeCodePath: "claude",
+			apiModelId: "claude-3-5-sonnet-20241022",
+			claudeCodeMaxOutputTokens: 16384,
+		}
+		const handlerWithMaxTokens = new ClaudeCodeHandler(options)
+
+		const systemPrompt = "You are a helpful assistant"
+		const messages = [{ role: "user" as const, content: "Hello" }]
+		const filteredMessages = [{ role: "user" as const, content: "Hello (filtered)" }]
+
+		mockFilterMessages.mockReturnValue(filteredMessages)
+
+		// Mock empty async generator
+		const mockGenerator = async function* (): AsyncGenerator<ClaudeCodeMessage | string> {
+			// Empty generator for basic test
+		}
+		mockRunClaudeCode.mockReturnValue(mockGenerator())
+
+		const stream = handlerWithMaxTokens.createMessage(systemPrompt, messages)
+
+		// Need to start iterating to trigger the call
+		const iterator = stream[Symbol.asyncIterator]()
+		await iterator.next()
+
+		// Verify runClaudeCode was called with maxOutputTokens
+		expect(mockRunClaudeCode).toHaveBeenCalledWith({
+			systemPrompt,
+			messages: filteredMessages,
+			path: "claude",
+			modelId: "claude-3-5-sonnet-20241022",
+			maxOutputTokens: 16384,
 		})
 	})
 

src/api/providers/claude-code.ts

@@ -1,5 +1,5 @@
 import type { Anthropic } from "@anthropic-ai/sdk"
-import { claudeCodeDefaultModelId, type ClaudeCodeModelId, claudeCodeModels } from "@roo-code/types"
+import { claudeCodeDefaultModelId, type ClaudeCodeModelId, claudeCodeModels, type ModelInfo } from "@roo-code/types"
 import { type ApiHandler } from ".."
 import { ApiStreamUsageChunk, type ApiStream } from "../transform/stream"
 import { runClaudeCode } from "../../integrations/claude-code/run"
@@ -25,6 +25,7 @@ export class ClaudeCodeHandler extends BaseProvider implements ApiHandler {
 			messages: filteredMessages,
 			path: this.options.claudeCodePath,
 			modelId: this.getModel().id,
+			maxOutputTokens: this.options.claudeCodeMaxOutputTokens,
 		})
 
 		// Usage is included with assistant messages,
@@ -129,12 +130,26 @@ export class ClaudeCodeHandler extends BaseProvider implements ApiHandler {
 		const modelId = this.options.apiModelId
 		if (modelId && modelId in claudeCodeModels) {
 			const id = modelId as ClaudeCodeModelId
-			return { id, info: claudeCodeModels[id] }
+			const modelInfo: ModelInfo = { ...claudeCodeModels[id] }
+
+			// Override maxTokens with the configured value if provided
+			if (this.options.claudeCodeMaxOutputTokens !== undefined) {
+				modelInfo.maxTokens = this.options.claudeCodeMaxOutputTokens
+			}
+
+			return { id, info: modelInfo }
+		}
+
+		const defaultModelInfo: ModelInfo = { ...claudeCodeModels[claudeCodeDefaultModelId] }
+
+		// Override maxTokens with the configured value if provided
+		if (this.options.claudeCodeMaxOutputTokens !== undefined) {
+			defaultModelInfo.maxTokens = this.options.claudeCodeMaxOutputTokens
 		}
 
 		return {
 			id: claudeCodeDefaultModelId,
-			info: claudeCodeModels[claudeCodeDefaultModelId],
+			info: defaultModelInfo,
 		}
 	}
 

src/integrations/claude-code/run.ts

@@ -3,6 +3,7 @@ import type Anthropic from "@anthropic-ai/sdk"
 import { execa } from "execa"
 import { ClaudeCodeMessage } from "./types"
 import readline from "readline"
+import { CLAUDE_CODE_DEFAULT_MAX_OUTPUT_TOKENS } from "@roo-code/types"
 
 const cwd = vscode.workspace.workspaceFolders?.map((folder) => folder.uri.fsPath).at(0)
 
@@ -20,7 +21,9 @@ type ProcessState = {
 	exitCode: number | null
 }
 
-export async function* runClaudeCode(options: ClaudeCodeOptions): AsyncGenerator<ClaudeCodeMessage | string> {
+export async function* runClaudeCode(
+	options: ClaudeCodeOptions & { maxOutputTokens?: number },
+): AsyncGenerator<ClaudeCodeMessage | string> {
 	const process = runProcess(options)
 
 	const rl = readline.createInterface({
@@ -107,7 +110,13 @@ const claudeCodeTools = [
 
 const CLAUDE_CODE_TIMEOUT = 600000 // 10 minutes
 
-function runProcess({ systemPrompt, messages, path, modelId }: ClaudeCodeOptions) {
+function runProcess({
+	systemPrompt,
+	messages,
+	path,
+	modelId,
+	maxOutputTokens,
+}: ClaudeCodeOptions & { maxOutputTokens?: number }) {
 	const claudePath = path || "claude"
 
 	const args = [
@@ -134,8 +143,11 @@ function runProcess({ systemPrompt, messages, path, modelId }: ClaudeCodeOptions
 		stderr: "pipe",
 		env: {
 			...process.env,
-			// The default is 32000. However, I've gotten larger responses, so we increase it unless the user specified it.
-			CLAUDE_CODE_MAX_OUTPUT_TOKENS: process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS || "64000",
+			// Use the configured value, or the environment variable, or default to CLAUDE_CODE_DEFAULT_MAX_OUTPUT_TOKENS
+			CLAUDE_CODE_MAX_OUTPUT_TOKENS:
+				maxOutputTokens?.toString() ||
+				process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS ||
+				CLAUDE_CODE_DEFAULT_MAX_OUTPUT_TOKENS.toString(),
 		},
 		cwd,
 		maxBuffer: 1024 * 1024 * 1000,