Merge pull request #104 from alibaba/refactor/mv-brain-outof-llms

refactor: move agent logic out of `llms`

packages/llms/src/OpenAIClient.ts

@@ -1,60 +1,58 @@
 /**
  * OpenAI Client implementation
- * @note This client is only for demonstrating how to implement a LLM client.
- * @note Use OpenAILenientClient instead.
  */
 import { InvokeError, InvokeErrorType } from './errors'
-import type { InvokeResult, LLMClient, LLMConfig, Message, Tool } from './types'
+import type { InvokeOptions, InvokeResult, LLMClient, LLMConfig, Message, Tool } from './types'
 import { modelPatch, zodToOpenAITool } from './utils'
 
 /**
- * @deprecated Use OpenAILenientClient instead.
+ * Client for OpenAI compatible APIs
  */
 export class OpenAIClient implements LLMClient {
-	config: LLMConfig
+	config: Required<LLMConfig>
+	private fetch: typeof globalThis.fetch
 
-	constructor(config: LLMConfig) {
+	constructor(config: Required<LLMConfig>) {
 		this.config = config
+		this.fetch = config.customFetch
 	}
 
 	async invoke(
 		messages: Message[],
 		tools: Record<string, Tool>,
-		abortSignal?: AbortSignal
+		abortSignal?: AbortSignal,
+		options?: InvokeOptions
 	): Promise<InvokeResult> {
 		// 1. Convert tools to OpenAI format
-		const openaiTools = Object.entries(tools).map(([name, tool]) => zodToOpenAITool(name, tool))
+		const openaiTools = Object.entries(tools).map(([name, t]) => zodToOpenAITool(name, t))
+
+		// Build request body
+		const requestBody: Record<string, unknown> = {
+			model: this.config.model,
+			temperature: this.config.temperature,
+			messages,
+			tools: openaiTools,
+			parallel_tool_calls: false,
+			// Require tool call: specific tool if provided, otherwise any tool
+			tool_choice: options?.toolChoiceName
+				? { type: 'function', function: { name: options.toolChoiceName } }
+				: 'required',
+		}
 
 		// 2. Call API
 		let response: Response
 		try {
-			response = await fetch(`${this.config.baseURL}/chat/completions`, {
+			response = await this.fetch(`${this.config.baseURL}/chat/completions`, {
 				method: 'POST',
 				headers: {
 					'Content-Type': 'application/json',
 					Authorization: `Bearer ${this.config.apiKey}`,
 				},
-				body: JSON.stringify(
+				body: JSON.stringify(modelPatch(requestBody)),
-					modelPatch({
-						model: this.config.model,
-						temperature: this.config.temperature,
-						messages,
-
-						tools: openaiTools,
-						// tool_choice: 'required',
-						tool_choice: { type: 'function', function: { name: 'AgentOutput' } },
-
-						// model specific params
-
-						// reasoning_effort: 'minimal',
-						// verbosity: 'low',
-						parallel_tool_calls: false,
-					})
-				),
 				signal: abortSignal,
 			})
 		} catch (error: unknown) {
-			// Network error
+			console.error(error)
 			throw new InvokeError(InvokeErrorType.NETWORK_ERROR, 'Network request failed', error)
 		}
 
@@ -92,77 +90,93 @@ export class OpenAIClient implements LLMClient {
 			)
 		}
 
+		// 4. Parse and validate response
 		const data = await response.json()
 
-		// 4. Check finish_reason
 		const choice = data.choices?.[0]
 		if (!choice) {
 			throw new InvokeError(InvokeErrorType.UNKNOWN, 'No choices in response', data)
 		}
 
+		// Check finish_reason
 		switch (choice.finish_reason) {
 			case 'tool_calls':
-				// ✅ Normal
+			case 'function_call': // gemini
+			case 'stop': // some models use this even with tool calls
 				break
 			case 'length':
-				// ⚠️ Token limit reached
 				throw new InvokeError(
 					InvokeErrorType.CONTEXT_LENGTH,
-					'Response truncated: max tokens reached',
+					'Response truncated: max tokens reached'
-					data
 				)
 			case 'content_filter':
-				// ❌ Content filtered
+				throw new InvokeError(InvokeErrorType.CONTENT_FILTER, 'Content filtered by safety system')
-				throw new InvokeError(
-					InvokeErrorType.CONTENT_FILTER,
-					'Content filtered by safety system',
-					data
-				)
-			case 'stop':
-				// ❌ Did not call tool (we require tool call)
-				throw new InvokeError(InvokeErrorType.NO_TOOL_CALL, 'Model did not call any tool', data)
 			default:
 				throw new InvokeError(
 					InvokeErrorType.UNKNOWN,
-					`Unexpected finish_reason: ${choice.finish_reason}`,
+					`Unexpected finish_reason: ${choice.finish_reason}`
-					data
 				)
 		}
 
-		// 5. Parse tool call
+		// Apply normalizeResponse if provided (for fixing format issues automatically)
-		const toolCall = choice.message?.tool_calls?.[0]
+		const normalizedData = options?.normalizeResponse ? options.normalizeResponse(data) : data
-		if (!toolCall) {
+		const normalizedChoice = (normalizedData as any).choices?.[0]
-			throw new InvokeError(InvokeErrorType.NO_TOOL_CALL, 'No tool call found in response', data)
-		}
 
-		const toolName = toolCall.function.name
+		// Get tool name from response
-		const tool = tools[toolName]
+		const toolCallName = normalizedChoice?.message?.tool_calls?.[0]?.function?.name
-		if (!tool) {
+		if (!toolCallName) {
-			throw new InvokeError(InvokeErrorType.UNKNOWN, `Tool ${toolName} not found`, data)
-		}
-
-		// 6. Parse and validate arguments
-		let toolArgs: unknown
-		try {
-			toolArgs = JSON.parse(toolCall.function.arguments)
-		} catch (e) {
-			throw new InvokeError(InvokeErrorType.INVALID_TOOL_ARGS, 'Invalid JSON in tool arguments', e)
-		}
-
-		// Validate against zod schema
-		const validation = tool.inputSchema.safeParse(toolArgs)
-		if (!validation.success) {
 			throw new InvokeError(
-				InvokeErrorType.INVALID_TOOL_ARGS,
+				InvokeErrorType.NO_TOOL_CALL,
-				`Tool arguments validation failed: ${validation.error.message}`,
+				'No tool call found in response',
-				validation.error
+				normalizedData
 			)
 		}
 
-		// 7. Execute tool
+		const tool = tools[toolCallName]
+		if (!tool) {
+			throw new InvokeError(
+				InvokeErrorType.UNKNOWN,
+				`Tool "${toolCallName}" not found in tools`,
+				normalizedData
+			)
+		}
+
+		// Extract and parse tool arguments
+		const argString = normalizedChoice.message?.tool_calls?.[0]?.function?.arguments
+		if (!argString) {
+			throw new InvokeError(
+				InvokeErrorType.INVALID_TOOL_ARGS,
+				'No tool call arguments found',
+				normalizedData
+			)
+		}
+
+		let parsedArgs: unknown
+		try {
+			parsedArgs = JSON.parse(argString)
+		} catch (error) {
+			throw new InvokeError(
+				InvokeErrorType.INVALID_TOOL_ARGS,
+				'Failed to parse tool arguments as JSON',
+				error
+			)
+		}
+
+		// Validate with schema
+		const validation = tool.inputSchema.safeParse(parsedArgs)
+		if (!validation.success) {
+			throw new InvokeError(
+				InvokeErrorType.INVALID_TOOL_ARGS,
+				'Tool arguments validation failed',
+				validation.error
+			)
+		}
+		const toolInput = validation.data
+
+		// 5. Execute tool
 		let toolResult: unknown
 		try {
-			toolResult = await tool.execute(validation.data)
+			toolResult = await tool.execute(toolInput)
 		} catch (e) {
 			throw new InvokeError(
 				InvokeErrorType.TOOL_EXECUTION_ERROR,
@@ -171,12 +185,11 @@ export class OpenAIClient implements LLMClient {
 			)
 		}
 
-		// 8. Return result (including cache tokens)
+		// Return result
 		return {
 			toolCall: {
-				// id: toolCall.id,
+				name: toolCallName,
-				name: toolName,
+				args: toolInput,
-				args: validation.data as Record<string, unknown>,
 			},
 			toolResult,
 			usage: {

packages/llms/src/OpenAILenientClient.ts

@@ -1,124 +0,0 @@
-/**
- * OpenAI Client implementation
- */
-import { InvokeError, InvokeErrorType } from './errors'
-import type { InvokeResult, LLMClient, LLMConfig, MacroToolInput, Message, Tool } from './types'
-import { lenientParseMacroToolCall, modelPatch, zodToOpenAITool } from './utils'
-
-export class OpenAIClient implements LLMClient {
-	config: Required<LLMConfig>
-	private fetch: typeof globalThis.fetch
-
-	constructor(config: Required<LLMConfig>) {
-		this.config = config
-		this.fetch = config.customFetch
-	}
-
-	async invoke(
-		messages: Message[],
-		tools: { AgentOutput: Tool<MacroToolInput> },
-		abortSignal?: AbortSignal
-	): Promise<InvokeResult> {
-		// 1. Convert tools to OpenAI format
-		const openaiTools = Object.entries(tools).map(([name, tool]) => zodToOpenAITool(name, tool))
-
-		// 2. Call API
-		let response: Response
-		try {
-			response = await this.fetch(`${this.config.baseURL}/chat/completions`, {
-				method: 'POST',
-				headers: {
-					'Content-Type': 'application/json',
-					Authorization: `Bearer ${this.config.apiKey}`,
-				},
-				body: JSON.stringify(
-					modelPatch({
-						model: this.config.model,
-						temperature: this.config.temperature,
-						messages,
-
-						tools: openaiTools,
-						// tool_choice: 'required',
-						tool_choice: { type: 'function', function: { name: 'AgentOutput' } },
-						parallel_tool_calls: false,
-					})
-				),
-				signal: abortSignal,
-			})
-		} catch (error: unknown) {
-			// Network error
-			console.error(error)
-			throw new InvokeError(InvokeErrorType.NETWORK_ERROR, 'Network request failed', error)
-		}
-
-		// 3. Handle HTTP errors
-		if (!response.ok) {
-			const errorData = await response.json().catch()
-			const errorMessage =
-				(errorData as { error?: { message?: string } }).error?.message || response.statusText
-
-			if (response.status === 401 || response.status === 403) {
-				throw new InvokeError(
-					InvokeErrorType.AUTH_ERROR,
-					`Authentication failed: ${errorMessage}`,
-					errorData
-				)
-			}
-			if (response.status === 429) {
-				throw new InvokeError(
-					InvokeErrorType.RATE_LIMIT,
-					`Rate limit exceeded: ${errorMessage}`,
-					errorData
-				)
-			}
-			if (response.status >= 500) {
-				throw new InvokeError(
-					InvokeErrorType.SERVER_ERROR,
-					`Server error: ${errorMessage}`,
-					errorData
-				)
-			}
-			throw new InvokeError(
-				InvokeErrorType.UNKNOWN,
-				`HTTP ${response.status}: ${errorMessage}`,
-				errorData
-			)
-		}
-
-		// parse response
-
-		const data = await response.json()
-		const tool = tools.AgentOutput
-		const macroToolInput = lenientParseMacroToolCall(data, tool.inputSchema as any)
-
-		// Execute tool
-		let toolResult: unknown
-		try {
-			toolResult = await tool.execute(macroToolInput)
-		} catch (e) {
-			throw new InvokeError(
-				InvokeErrorType.TOOL_EXECUTION_ERROR,
-				`Tool execution failed: ${(e as Error).message}`,
-				e
-			)
-		}
-
-		// Return result (including cache tokens)
-		return {
-			toolCall: {
-				// id: toolCall.id,
-				name: 'AgentOutput',
-				args: macroToolInput,
-			},
-			toolResult,
-			usage: {
-				promptTokens: data.usage?.prompt_tokens ?? 0,
-				completionTokens: data.usage?.completion_tokens ?? 0,
-				totalTokens: data.usage?.total_tokens ?? 0,
-				cachedTokens: data.usage?.prompt_tokens_details?.cached_tokens,
-				reasoningTokens: data.usage?.completion_tokens_details?.reasoning_tokens,
-			},
-			rawResponse: data,
-		}
-	}
-}

packages/llms/src/index.ts

@@ -1,37 +1,4 @@
-/**
+import { OpenAIClient } from './OpenAIClient'
- * @topic LLM 与主流程的隔离
- * @reasoning
- * 将 llm 的调用和主流程分开是复杂的，
- * 因为 agent 的 tool call 通常集成在 llm 模块中，而而先得到 llm 返回，然后处理工具调用
- * tools 和 llm 调用的逻辑不可避免地耦合在一起，tool 的执行又和主流程耦合在一起
- * 而 history 的维护和更新逻辑，又必须嵌入多轮 tool call 中
- * @reasoning
- * - 放弃框架提供的自动的多轮调用，每轮调用都由主流程发起
- * - 理想情况下，llm 调用应该获得 structured output，然后由额外的模块触发 tool call，目前模型和框架都无法实现
- * - 当前只能将 llm api 和 本地 tool call 耦合在一起，不关心其中的衔接方式
- * @conclusion
- * - @llm responsibility boundary:
- *   - call llm api with given messages and tools
- *   - invoke tool call and get the result of the tool
- *   - return the result to main loop
- * - @main_loop responsibility boundary:
- *   - maintain all behaviors of an **agent**
- * @conclusion
- * - 这里的 llm 模块不是 agent，只负责一轮 llm 调用和工具调用，无状态
- */
-/**
- * @topic 结构化输出
- * @facts
- * - 几乎所有模型都支持 tool call schema
- * - 几乎所有模型都支持返回 json
- *   - 只有 openAI/grok/gemini 支持 schema 并保证格式
- * - 主流模型都支持 tool_choice: required
- *   - 除了 qwen 必须指定一个函数名 (9月上新后支持)
- * @conclusion
- * - 永远使用 tool call 来返回结构化数据，禁止模型直接返回（视为出错）
- * - 不能假设 tool 参数合法，必须有修复机制，而且修复也应该使用 tool call 返回
- */
-import { OpenAIClient } from './OpenAILenientClient'
 import {
 	DEFAULT_API_KEY,
 	DEFAULT_BASE_URL,
@@ -40,27 +7,9 @@ import {
 	LLM_MAX_RETRIES,
 } from './constants'
 import { InvokeError } from './errors'
-import type {
+import type { InvokeOptions, InvokeResult, LLMClient, LLMConfig, Message, Tool } from './types'
-	AgentBrain,
-	InvokeResult,
-	LLMClient,
-	LLMConfig,
-	MacroToolInput,
-	MacroToolResult,
-	Message,
-	Tool,
-} from './types'
 
-export type {
+export type { InvokeOptions, InvokeResult, LLMClient, LLMConfig, Message, Tool }
-	AgentBrain,
-	InvokeResult,
-	LLMClient,
-	LLMConfig,
-	MacroToolInput,
-	MacroToolResult,
-	Message,
-	Tool,
-}
 
 export function parseLLMConfig(config: LLMConfig): Required<LLMConfig> {
 	return {
@@ -93,11 +42,12 @@ export class LLM extends EventTarget {
 	async invoke(
 		messages: Message[],
 		tools: Record<string, Tool>,
-		abortSignal: AbortSignal
+		abortSignal: AbortSignal,
+		options?: InvokeOptions
 	): Promise<InvokeResult> {
 		return await withRetry(
 			async () => {
-				const result = await this.client.invoke(messages, tools, abortSignal)
+				const result = await this.client.invoke(messages, tools, abortSignal, options)
 
 				return result
 			},

packages/llms/src/types.ts

@@ -32,6 +32,24 @@ export interface Tool<TParams = any, TResult = any> {
 	execute: (args: TParams) => Promise<TResult>
 }
 
+/**
+ * Invoke options for LLM call
+ */
+export interface InvokeOptions {
+	/**
+	 * Force LLM to call a specific tool by name.
+	 * If provided: tool_choice = { type: 'function', function: { name: toolChoiceName } }
+	 * If not provided: tool_choice = 'required' (must call some tool, but model chooses which)
+	 */
+	toolChoiceName?: string
+	/**
+	 * Response normalization function.
+	 * Called before parsing the response.
+	 * Used to fix various response format errors from the model.
+	 */
+	normalizeResponse?: (response: any) => any
+}
+
 /**
  * LLM Client interface
  * Note: Does not use generics because each tool in the tools array has different types
@@ -40,7 +58,8 @@ export interface LLMClient {
 	invoke(
 		messages: Message[],
 		tools: Record<string, Tool>,
-		abortSignal?: AbortSignal
+		abortSignal?: AbortSignal,
+		options?: InvokeOptions
 	): Promise<InvokeResult>
 }
 
@@ -82,36 +101,3 @@ export interface LLMConfig {
 	 */
 	customFetch?: typeof globalThis.fetch
 }
-
-/**
- * Agent brain state - the reflection-before-action model
- *
- * Every tool call must first reflect on:
- * - evaluation_previous_goal: How well did the previous action achieve its goal?
- * - memory: Key information to remember for future steps
- * - next_goal: What should be accomplished in the next action?
- */
-export interface AgentBrain {
-	// thinking?: string
-	evaluation_previous_goal: string
-	memory: string
-	next_goal: string
-}
-
-/**
- * MacroTool input structure
- *
- * This is the core abstraction that enforces the "reflection-before-action" mental model.
- * Before executing any action, the LLM must output its reasoning state.
- */
-export interface MacroToolInput extends AgentBrain {
-	action: Record<string, any>
-}
-
-/**
- * MacroTool output structure
- */
-export interface MacroToolResult {
-	input: MacroToolInput
-	output: string
-}

packages/llms/src/utils.ts

@@ -4,8 +4,7 @@
 import chalk from 'chalk'
 import { z } from 'zod'
 
-import { InvokeError, InvokeErrorType } from './errors'
+import type { Tool } from './types'
-import type { MacroToolInput, Tool } from './types'
 
 function debug(message: string) {
 	console.debug(chalk.gray('[LLM]'), message)
@@ -26,176 +25,6 @@ export function zodToOpenAITool(name: string, tool: Tool) {
 	}
 }
 
-/**
- * Although some models cannot guarantee correct response. Common issues are fixable:
- * - Instead of returning a proper tool call. Return the tool call parameters in the message content.
- * - Returned tool calls or messages don't follow the nested MacroToolInput format.
- */
-export function lenientParseMacroToolCall(
-	responseData: any,
-	inputSchema: z.ZodObject<MacroToolInput & Record<string, any>>
-): MacroToolInput {
-	// check
-	const choice = responseData.choices?.[0]
-	if (!choice) {
-		throw new InvokeError(InvokeErrorType.UNKNOWN, 'No choices in response', responseData)
-	}
-
-	// check
-	switch (choice.finish_reason) {
-		case 'tool_calls':
-		case 'function_call': // gemini
-		case 'stop': // will try a robust parse
-			// ✅ Normal
-			break
-		case 'length':
-			// ⚠️ Token limit reached
-			throw new InvokeError(
-				InvokeErrorType.CONTEXT_LENGTH,
-				'Response truncated: max tokens reached'
-			)
-		case 'content_filter':
-			// ❌ Content filtered
-			throw new InvokeError(InvokeErrorType.CONTENT_FILTER, 'Content filtered by safety system')
-		default:
-			throw new InvokeError(
-				InvokeErrorType.UNKNOWN,
-				`Unexpected finish_reason: ${choice.finish_reason}`
-			)
-	}
-
-	// Extract action schema from MacroToolInput schema
-	const actionSchema = inputSchema.shape.action
-	if (!actionSchema) {
-		throw new Error('inputSchema must have an "action" field')
-	}
-
-	// patch stopReason mis-format
-
-	let arg: string | null = null
-
-	// try to use tool call
-	const toolCall = choice.message?.tool_calls?.[0]?.function
-	arg = toolCall?.arguments ?? null
-
-	if (arg && toolCall.name !== 'AgentOutput') {
-		// TODO: check if toolCall.name is a valid action name
-		// case: instead of AgentOutput, the model returned a action name as tool call
-		console.log(chalk.yellow('lenientParseMacroToolCall: #1 fixing incorrect tool call'))
-		let tmpArg
-		try {
-			tmpArg = JSON.parse(arg)
-		} catch (error) {
-			throw new InvokeError(
-				InvokeErrorType.INVALID_TOOL_ARGS,
-				'Failed to parse tool arguments as JSON',
-				error
-			)
-		}
-		arg = JSON.stringify({ action: { [toolCall.name]: tmpArg } })
-	}
-
-	if (!arg) {
-		// try to use message content as JSON
-		arg = choice.message?.content.trim() || null
-	}
-
-	if (!arg) {
-		throw new InvokeError(
-			InvokeErrorType.NO_TOOL_CALL,
-			'No tool call or content found in response',
-			responseData
-		)
-	}
-
-	// make sure is valid JSON
-
-	let parsedArgs: any
-	try {
-		parsedArgs = JSON.parse(arg)
-	} catch (error) {
-		throw new InvokeError(
-			InvokeErrorType.INVALID_TOOL_ARGS,
-			'Failed to parse tool arguments as JSON',
-			error
-		)
-	}
-
-	// patch incomplete formats
-
-	if (parsedArgs.action || parsedArgs.evaluation_previous_goal || parsedArgs.next_goal) {
-		// case: nested MacroToolInput format (correct format)
-
-		// some models may give a empty action (they may think reasoning and action should be separate)
-		if (!parsedArgs.action) {
-			console.log(chalk.yellow('lenientParseMacroToolCall: #2 fixing incorrect tool call'))
-			parsedArgs.action = {
-				wait: { seconds: 1 },
-			}
-		}
-	} else if (parsedArgs.type && parsedArgs.function) {
-		// case: upper level function call format provided. only keep its arguments
-		// TODO: check if function name is a valid action name
-		if (parsedArgs.function.name !== 'AgentOutput')
-			throw new InvokeError(
-				InvokeErrorType.INVALID_TOOL_ARGS,
-				`Expected function name "AgentOutput", got "${parsedArgs.function.name}"`,
-				null
-			)
-
-		console.log(chalk.yellow('lenientParseMacroToolCall: #3 fixing incorrect tool call'))
-		parsedArgs = parsedArgs.function.arguments
-	} else if (parsedArgs.name && parsedArgs.arguments) {
-		// case: upper level function call format provided. only keep its arguments
-		// TODO: check if function name is a valid action name
-		if (parsedArgs.name !== 'AgentOutput')
-			throw new InvokeError(
-				InvokeErrorType.INVALID_TOOL_ARGS,
-				`Expected function name "AgentOutput", got "${parsedArgs.name}"`,
-				null
-			)
-
-		console.log(chalk.yellow('lenientParseMacroToolCall: #4 fixing incorrect tool call'))
-		parsedArgs = parsedArgs.arguments
-	} else {
-		// case: only action parameters provided, wrap into MacroToolInput
-		// TODO: check if action name is valid
-		console.log(chalk.yellow('lenientParseMacroToolCall: #5 fixing incorrect tool call'))
-		parsedArgs = { action: parsedArgs } as MacroToolInput
-	}
-
-	// make sure it's not wrapped as string
-	if (typeof parsedArgs === 'string') {
-		console.log(chalk.yellow('lenientParseMacroToolCall: #6 fixing incorrect tool call'))
-		try {
-			parsedArgs = JSON.parse(parsedArgs)
-		} catch (error) {
-			throw new InvokeError(
-				InvokeErrorType.INVALID_TOOL_ARGS,
-				'Failed to parse nested tool arguments as JSON',
-				error
-			)
-		}
-	}
-
-	const validation = inputSchema.safeParse(parsedArgs)
-	if (validation.success) {
-		return validation.data as unknown as MacroToolInput
-	} else {
-		const action = parsedArgs.action ?? {}
-		const actionName = Object.keys(action)[0] || 'unknown'
-		const actionArgs = JSON.stringify(action[actionName] || 'unknown')
-
-		// TODO: check if action name is valid. give a readable error message
-
-		throw new InvokeError(
-			InvokeErrorType.INVALID_TOOL_ARGS,
-			`Tool arguments validation failed: action "${actionName}" with args ${actionArgs}`,
-			validation.error
-		)
-	}
-}
-
 /**
  * Patch model specific parameters
  */
@@ -206,10 +35,19 @@ export function modelPatch(body: Record<string, any>) {
 	const modelName = normalizeModelName(model)
 
 	if (modelName.startsWith('claude')) {
-		debug('Applying Claude patch: change tool_choice and disable thinking')
+		debug('Applying Claude patch: disable thinking')
-		body.tool_choice = { type: 'tool', name: 'AgentOutput' }
 		body.thinking = { type: 'disabled' }
-		// body.reasoning = { enabled: 'disabled' }
+
+		// Convert tool_choice to Claude format
+		if (body.tool_choice === 'required') {
+			// 'required' -> { type: 'any' } (must call some tool)
+			debug('Applying Claude patch: convert tool_choice "required" to { type: "any" }')
+			body.tool_choice = { type: 'any' }
+		} else if (body.tool_choice?.function?.name) {
+			// { type: 'function', function: { name: '...' } } -> { type: 'tool', name: '...' }
+			debug('Applying Claude patch: convert tool_choice format')
+			body.tool_choice = { type: 'tool', name: body.tool_choice.function.name }
+		}
 	}
 
 	if (modelName.startsWith('grok')) {

packages/page-agent/src/PageAgent.ts

@@ -2,13 +2,7 @@
  * Copyright (C) 2025 Alibaba Group Holding Limited
  * All rights reserved.
  */
-import {
+import { LLM, type Tool } from '@page-agent/llms'
-	type AgentBrain,
-	LLM,
-	type MacroToolInput,
-	type MacroToolResult,
-	type Tool,
-} from '@page-agent/llms'
 import { PageController } from '@page-agent/page-controller'
 import { Panel, SimulatorMask } from '@page-agent/ui'
 import chalk from 'chalk'
@@ -18,15 +12,46 @@ import type { PageAgentConfig } from './config'
 import { MAX_STEPS } from './config/constants'
 import SYSTEM_PROMPT from './prompts/system_prompt.md?raw'
 import { tools } from './tools'
-import { trimLines, uid, waitUntil } from './utils'
+import { normalizeResponse, trimLines, uid, waitUntil } from './utils'
 import { assert } from './utils/assert'
 
+/**
+ * Agent brain state - the reflection-before-action model
+ *
+ * Every tool call must first reflect on:
+ * - evaluation_previous_goal: How well did the previous action achieve its goal?
+ * - memory: Key information to remember for future steps
+ * - next_goal: What should be accomplished in the next action?
+ */
+export interface AgentReflection {
+	evaluation_previous_goal: string
+	memory: string
+	next_goal: string
+}
+
+/**
+ * MacroTool input structure
+ *
+ * This is the core abstraction that enforces the "reflection-before-action" mental model.
+ * Before executing any action, the LLM must output its reasoning state.
+ */
+export interface MacroToolInput extends Partial<AgentReflection> {
+	action: Record<string, any>
+}
+
+/**
+ * MacroTool output structure
+ */
+export interface MacroToolResult {
+	input: MacroToolInput
+	output: string
+}
+
 export type { PageAgentConfig }
 export { tool, type PageAgentTool } from './tools'
-export type { AgentBrain, MacroToolInput, MacroToolResult }
 
 export interface AgentHistory {
-	brain: AgentBrain
+	brain: Partial<AgentReflection>
 	action: {
 		name: string
 		input: any
@@ -124,9 +149,6 @@ export class PageAgent extends EventTarget {
 		window.addEventListener('beforeunload', this.#beforeUnloadListener)
 	}
 
-	/**
-	 * @todo maybe return something?
-	 */
 	async execute(task: string): Promise<ExecutionResult> {
 		if (!task) throw new Error('Task is required')
 		this.task = task
@@ -183,7 +205,11 @@ export class PageAgent extends EventTarget {
 						},
 					],
 					{ AgentOutput: this.#packMacroTool() },
-					this.#abortController.signal
+					this.#abortController.signal,
+					{
+						toolChoiceName: 'AgentOutput',
+						normalizeResponse,
+					}
 				)
 
 				const macroResult = result.toolResult as MacroToolResult

packages/page-agent/src/utils/autoFixer.ts

@@ -0,0 +1,154 @@
+import chalk from 'chalk'
+
+/**
+ * Normalize LLM response and fix common format issues.
+ *
+ * Handles:
+ * - No tool_calls but JSON in message.content (fallback)
+ * - Model returns action name as tool call instead of AgentOutput
+ * - Arguments wrapped as double JSON string
+ * - Nested function call format
+ * - Missing action field (fallback to wait)
+ * - etc.
+ */
+export function normalizeResponse(response: any): any {
+	let resolvedArguments = null as any
+
+	const choice = (response as { choices?: Choice[] }).choices?.[0]
+	if (!choice) throw new Error('No choices in response')
+
+	const message = choice.message
+	if (!message) throw new Error('No message in choice')
+
+	const toolCall = message.tool_calls?.[0]
+
+	// fix level and location of arguments
+
+	if (toolCall?.function?.arguments) {
+		resolvedArguments = safeJsonParse(toolCall.function.arguments)
+
+		// case: sometimes the model only returns the action level
+		if (toolCall.function.name && toolCall.function.name !== 'AgentOutput') {
+			console.log(chalk.yellow(`[normalizeResponse] #1: fixing tool_call`))
+			resolvedArguments = { action: safeJsonParse(resolvedArguments) }
+		}
+	} else {
+		// case: sometimes the model returns json in content instead of tool_calls
+		if (message.content) {
+			const content = message.content.trim()
+			const jsonInContent = retrieveJsonFromString(content)
+			if (jsonInContent) {
+				resolvedArguments = safeJsonParse(jsonInContent)
+
+				// case: sometimes the content json includes upper level wrapper
+				if (resolvedArguments?.name === 'AgentOutput') {
+					console.log(chalk.yellow(`[normalizeResponse] #2: fixing tool_call`))
+					resolvedArguments = safeJsonParse(resolvedArguments.arguments)
+				}
+
+				// case: sometimes even 2-levels of wrapping
+				if (resolvedArguments?.type === 'function') {
+					console.log(chalk.yellow(`[normalizeResponse] #3: fixing tool_call`))
+					resolvedArguments = safeJsonParse(resolvedArguments.function.arguments)
+				}
+
+				// case: and sometimes action level only
+				// todo: needs better detection logic
+				if (
+					!resolvedArguments?.action &&
+					!resolvedArguments?.evaluation_previous_goal &&
+					!resolvedArguments?.memory &&
+					!resolvedArguments?.next_goal &&
+					!resolvedArguments?.thinking
+				) {
+					console.log(chalk.yellow(`[normalizeResponse] #4: fixing tool_call`))
+					resolvedArguments = { action: safeJsonParse(resolvedArguments) }
+				}
+			} else {
+				throw new Error('No tool_call and the message content does not contain valid JSON')
+			}
+		} else {
+			throw new Error('No tool_call nor message content is present')
+		}
+	}
+
+	// fix double stringified arguments
+	resolvedArguments = safeJsonParse(resolvedArguments)
+
+	// fix incomplete formats
+	if (!resolvedArguments.action) {
+		console.log(chalk.yellow(`[normalizeResponse] #5: fixing tool_call`))
+		resolvedArguments.action = { name: 'wait', input: { seconds: 1 } }
+	}
+
+	// pack back to standard format
+	return {
+		...response,
+		choices: [
+			{
+				...choice,
+				message: {
+					...message,
+					tool_calls: [
+						{
+							...(toolCall || {}),
+							function: {
+								...(toolCall?.function || {}),
+								name: 'AgentOutput',
+								arguments: JSON.stringify(resolvedArguments),
+							},
+						},
+					],
+				},
+			},
+		],
+	}
+}
+
+/**
+ * Safely parse JSON, return original input if not json.
+ */
+function safeJsonParse(input: any): any {
+	if (typeof input === 'string') {
+		try {
+			return JSON.parse(input.trim())
+		} catch {
+			return input
+		}
+	}
+	return input
+}
+
+/**
+ * Extract and parse JSON from a string.
+ * - Treat content between the first `{` and the last `}` as JSON.
+ * - Try to parse that content as JSON and return the parsed value (object/array/primitive) if successful, otherwise return null.
+ */
+function retrieveJsonFromString(str: string): any {
+	try {
+		const json = /({[\s\S]*})/.exec(str) ?? []
+		if (json.length === 0) {
+			return null
+		}
+		return JSON.parse(json[0]!)
+	} catch {
+		return null
+	}
+}
+
+interface Choice {
+	message?: {
+		role?: 'assistant'
+		content?: string
+		tool_calls?: {
+			id?: string
+			type?: 'function'
+			function?: {
+				name?: string
+				arguments?: string
+			}
+		}[]
+	}
+	index?: 0
+	finish_reason?: 'tool_calls'
+}

packages/page-agent/src/utils/index.ts

@@ -1,3 +1,5 @@
+export { normalizeResponse } from './autoFixer'
+
 /**
  * Wait until condition becomes true
  * @returns Returns when condition becomes true, throws otherwise