Merge pull request #119 from alibaba/feat/decouple-dom

feat!: decouple `DOM` from `Agent`
2026-01-19 17:22:57 +08:00
parent 8abe3afed9 0643cc4ff9
commit 8bde218f28
48 changed files with 10695 additions and 9745 deletions
--- a/.vscode/settings.json
+++ b/.vscode/settings.json
@@ -2,6 +2,7 @@
    "editor.fontLigatures": true,
    "cSpell.words": [
        "deepseek",
+        "historychange",
        "HITL",
        "innerhtml",
        "llms",
@@ -10,6 +11,7 @@
        "qwen",
        "retryable",
        "shadcn",
+        "statuschange",
        "wouter"
    ],
    "markdownlint.config": {
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -4,11 +4,12 @@

 This is a **monorepo** with npm workspaces:

- **Core Library** (`packages/page-agent/`) - AI agent for browser DOM automation, published as `page-agent` on npm
+- **Page Agent** (`packages/page-agent/`) - Main entry with built-in UI Panel, published as `page-agent` on npm
 - **Website** (`packages/website/`) - React docs and landing page. **When working on website, follow `packages/website/AGENTS.md`**

 Internal packages:

+- **Core** (`packages/core/`) - PageAgentCore without UI (npm: `@page-agent/core`)
 - **CDN** (`packages/cdn/`) - IIFE builds for script tag usage (npm: `@page-agent/cdn`)
 - **LLMs** (`packages/llms/`) - LLM client with reflection-before-action mental model
 - **Page Controller** (`packages/page-controller/`) - DOM operations and visual feedback (SimulatorMask), independent of LLM
@@ -31,7 +32,8 @@ Simple monorepo solution: TypeScript references + Vite aliases. Update tsconfig

 ```
 packages/
-├── page-agent/              # npm: "page-agent" ⭐ MAIN
+├── page-agent/              # npm: "page-agent" ⭐ MAIN (with Panel UI)
+├── core/                    # npm: "@page-agent/core" (headless, no UI)
 ├── cdn/                     # npm: "@page-agent/cdn" (IIFE builds)
 ├── website/                 # @page-agent/website (private)
 ├── llms/                    # @page-agent/llms
@@ -43,9 +45,10 @@ packages/

 ### Module Boundaries

- **Page Agent**: Core lib. Imports from `@page-agent/llms`, `@page-agent/page-controller`, `@page-agent/ui`
+- **Page Agent**: Main entry with UI. Extends PageAgentCore and adds Panel. Imports from `@page-agent/core`, `@page-agent/ui`
+- **Core**: PageAgentCore without UI. Imports from `@page-agent/llms`, `@page-agent/page-controller`
 - **LLMs**: LLM client with MacroToolInput contract. No dependency on page-agent
- **UI**: Panel and i18n. No dependency on page-agent
+- **UI**: Panel and i18n. Decoupled from PageAgent via PanelAgentAdapter interface
 - **Page Controller**: DOM operations with optional visual feedback (SimulatorMask). No LLM dependency. Enable mask via `enableMask: true` config

 ### PageController ↔ PageAgent Communication
@@ -87,10 +90,18 @@ Demo build supports query params (e.g., `?model=gpt-4&lang=en-US`).
 ### Page Agent (`packages/page-agent/`)

 | File               | Description                                    |
-| ------------------ | --------------------------------------- |
-| `src/PageAgent.ts` | ⭐ Main AI agent class                  |
-| `src/umd.ts`       | CDN/UMD entry with auto-init            |
+| ------------------ | ---------------------------------------------- |
+| `src/PageAgent.ts` | ⭐ Main class with UI, extends PageAgentCore   |
+| `src/iife.ts`      | IIFE/CDN entry                                 |
+
+### Core (`packages/core/`)
+
+| File                    | Description                                 |
+| ----------------------- | ------------------------------------------- |
+| `src/PageAgentCore.ts`  | ⭐ Core agent class without UI              |
 | `src/tools/`            | Tool definitions calling PageController     |
+| `src/config/`           | Configuration types and constants           |
+| `src/prompts/`          | System prompt templates                     |

 ### LLMs (`packages/llms/`)

@@ -113,7 +124,7 @@ Demo build supports query params (e.g., `?model=gpt-4&lang=en-US`).

 ### New Agent Tool

-1. Implement in `packages/page-agent/src/tools/index.ts`
+1. Implement in `packages/core/src/tools/index.ts`
 2. If tool needs DOM ops, add method to PageController first
 3. Tool calls `this.pageController.methodName()` for DOM interactions

--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -20,10 +20,11 @@ Thank you for your interest in contributing to Page-Agent! We welcome contributi

 ### Project Structure

-This is a **monorepo** with npm workspaces containing **two main packages**:
+This is a **monorepo** with npm workspaces containing **3 main packages**:

-1. **Core Library** (`packages/page-agent/`) - Pure JavaScript/TypeScript AI agent library for browser DOM automation, published as `page-agent` on npm
-2. **Website** (`packages/website/`) - React documentation and landing page. Also as demo and test page for the core lib. private package `@page-agent/website`
+- **Page Agent** (`packages/page-agent/`) - Main entry with built-in UI Panel, published as `page-agent` on npm
+- **Core** (`packages/core/`) - Core agent logic without UI (npm: `@page-agent/core`)
+- **Website** (`packages/website/`) - React documentation and landing page. Also as demo and test page for the core lib. private package `@page-agent/website`

 We use a simplified monorepo solution with `native npm-workspace + ts reference + vite alias`. No fancy tooling. Hoisting is required.

--- a/README-zh.md
+++ b/README-zh.md
@@ -76,7 +76,8 @@ PageAgent adopts a simplified monorepo structure:

 ```
 packages/
-├── page-agent/          # AI agent (npm: page-agent)
+├── page-agent/          # AI agent with UI Panel(npm: page-agent)
+├── core/                # Agent core logic without UI(npm: @page-agent/core)
 ├── llms/                # LLM 客户端 (npm: @page-agent/llms)
 ├── page-controller/     # DOM 操作 & 蒙层 & 模拟鼠标 (npm: @page-agent/page-controller)
 ├── ui/                  # 面板 & i18n (npm: @page-agent/ui)
--- a/README.md
+++ b/README.md
@@ -76,7 +76,8 @@ PageAgent adopts a simplified monorepo structure:

 ```
 packages/
-├── page-agent/          # AI agent (npm: page-agent)
+├── page-agent/          # AI agent with UI Panel(npm: page-agent)
+├── core/                # Agent core logic without UI(npm: @page-agent/core)
 ├── llms/                # LLM client (npm: @page-agent/llms)
 ├── page-controller/     # DOM operations & Visual Mask (npm: @page-agent/page-controller)
 ├── ui/                  # Panel & i18n (npm: @page-agent/ui)
--- a/package-lock.json
+++ b/package-lock.json
@@ -12,6 +12,7 @@
                "packages/page-controller",
                "packages/ui",
                "packages/llms",
+                "packages/core",
                "packages/page-agent",
                "packages/cdn",
                "packages/website"
@@ -1588,6 +1589,10 @@
            "resolved": "packages/cdn",
            "link": true
        },
+        "node_modules/@page-agent/core": {
+            "resolved": "packages/core",
+            "link": true
+        },
        "node_modules/@page-agent/llms": {
            "resolved": "packages/llms",
            "link": true
@@ -8139,6 +8144,17 @@
                "page-agent": "0.2.5"
            }
        },
+        "packages/core": {
+            "name": "@page-agent/core",
+            "version": "0.2.5",
+            "license": "MIT",
+            "dependencies": {
+                "@page-agent/llms": "0.2.5",
+                "@page-agent/page-controller": "0.2.5",
+                "chalk": "^5.6.2",
+                "zod": "^4.3.5"
+            }
+        },
        "packages/llms": {
            "name": "@page-agent/llms",
            "version": "0.2.5",
@@ -8152,6 +8168,7 @@
            "version": "0.2.5",
            "license": "MIT",
            "dependencies": {
+                "@page-agent/core": "0.2.5",
                "@page-agent/llms": "0.2.5",
                "@page-agent/page-controller": "0.2.5",
                "@page-agent/ui": "0.2.5",
--- a/package.json
+++ b/package.json
@@ -7,6 +7,7 @@
        "packages/page-controller",
        "packages/ui",
        "packages/llms",
+        "packages/core",
        "packages/page-agent",
        "packages/cdn",
        "packages/website"
--- a/packages/core/package.json
+++ b/packages/core/package.json
@@ -0,0 +1,51 @@
+{
+    "name": "@page-agent/core",
+    "private": false,
+    "version": "0.2.5",
+    "type": "module",
+    "main": "./dist/esm/page-agent-core.js",
+    "module": "./dist/esm/page-agent-core.js",
+    "types": "./dist/esm/PageAgentCore.d.ts",
+    "exports": {
+        ".": {
+            "types": "./dist/esm/PageAgentCore.d.ts",
+            "import": "./dist/esm/page-agent-core.js",
+            "default": "./dist/esm/page-agent-core.js"
+        }
+    },
+    "files": [
+        "dist/"
+    ],
+    "description": "GUI agent for web applications - add intelligent automation to any webpage with a single script",
+    "keywords": [
+        "ai",
+        "automation",
+        "ui-agent",
+        "GUI-agent",
+        "browser-automation",
+        "web-agent",
+        "llm",
+        "dom-interaction",
+        "web-automation",
+        "GUI-simulation"
+    ],
+    "author": "Simon<gaomeng1900>",
+    "license": "MIT",
+    "repository": {
+        "type": "git",
+        "url": "https://github.com/alibaba/page-agent.git"
+    },
+    "homepage": "https://alibaba.github.io/page-agent/",
+    "scripts": {
+        "build": "vite build",
+        "dev:iife": "concurrently \"vite build --config vite.iife.config.js --watch\" \"npx serve dist/iife -p 5174\"",
+        "prepublishOnly": "node -e \"const fs=require('fs');['README.md','LICENSE'].forEach(f=>fs.copyFileSync('../../'+f,f))\"",
+        "postpublish": "node -e \"['README.md','LICENSE'].forEach(f=>{try{require('fs').unlinkSync(f)}catch{}})\""
+    },
+    "dependencies": {
+        "chalk": "^5.6.2",
+        "zod": "^4.3.5",
+        "@page-agent/llms": "0.2.5",
+        "@page-agent/page-controller": "0.2.5"
+    }
+}
--- a/packages/core/src/PageAgentCore.ts
+++ b/packages/core/src/PageAgentCore.ts
@@ -0,0 +1,588 @@
+/**
+ * Copyright (C) 2025 Alibaba Group Holding Limited
+ * All rights reserved.
+ */
+import { LLM, type Tool } from '@page-agent/llms'
+import type { PageController } from '@page-agent/page-controller'
+import chalk from 'chalk'
+import zod from 'zod'
+
+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 {
+	AgentActivity,
+	AgentReflection,
+	AgentStatus,
+	AgentStep,
+	ExecutionResult,
+	HistoricalEvent,
+	MacroToolInput,
+	MacroToolResult,
+} from './types'
+import { normalizeResponse, trimLines, uid } from './utils'
+import { assert } from './utils/assert'
+
+export { type PageAgentConfig }
+export { tool, type PageAgentTool } from './tools'
+
+/**
+ * AI agent for browser DOM automation.
+ *
+ * @remarks
+ * ## Event System
+ * - `statuschange` - Agent status transitions (idle → running → completed/error)
+ * - `historychange` - History events updated (persistent, part of agent memory)
+ * - `activity` - Real-time activity feedback (transient, for UI only)
+ * - `dispose` - Agent cleanup triggered
+ *
+ * ## Information Streams
+ * 1. **History Events** (`history` array)
+ *    - Persistent event stream that forms agent's memory
+ *    - Included in LLM context across steps
+ *    - Types: steps, observations, user takeovers, llm errors
+ *
+ * 2. **Activity Events** (via `activity` event)
+ *    - Transient UI feedback during task execution
+ *    - NOT included in LLM context
+ *    - Types: thinking, executing, executed, retrying, error
+ */
+export class PageAgentCore extends EventTarget {
+	config: PageAgentConfig
+	id = uid()
+	tools: typeof tools
+	disposed = false
+	task = ''
+	taskId = ''
+
+	/** Agent execution status */
+	#status: AgentStatus = 'idle'
+
+	/**
+	 * Callback for when agent needs user input (ask_user tool)
+	 * If not set, ask_user tool will be disabled
+	 * @example onAskUser: (q) => window.prompt(q) || ''
+	 */
+	onAskUser?: (question: string) => Promise<string>
+
+	#llm: LLM
+	#abortController = new AbortController()
+
+	/** PageController for DOM operations */
+	pageController: PageController
+
+	/** Runtime states for tracking across steps */
+	states = {
+		/** Accumulated wait time in seconds, used by wait tool */
+		totalWaitTime: 0,
+		/** Last known URL for detecting navigation */
+		lastURL: '',
+	}
+
+	/** History events */
+	history: HistoricalEvent[] = []
+
+	constructor(config: PageAgentConfig & { pageController: PageController }) {
+		super()
+
+		this.config = config
+		this.#llm = new LLM(this.config)
+		this.tools = new Map(tools)
+		this.pageController = config.pageController
+
+		// Listen to LLM retry events
+		this.#llm.addEventListener('retry', (e) => {
+			const { attempt, maxAttempts } = (e as CustomEvent).detail
+			this.emitActivity({ type: 'retrying', attempt, maxAttempts })
+			// Also push to history for panel rendering
+			this.history.push({
+				type: 'error',
+				errorType: 'retry',
+				message: `LLM retry attempt ${attempt} of ${maxAttempts}`,
+				attempt,
+				maxAttempts,
+			})
+			this.#emitHistoryChange()
+		})
+		this.#llm.addEventListener('error', (e) => {
+			const { error } = (e as CustomEvent).detail
+			const message = String(error)
+			this.emitActivity({ type: 'error', message })
+			// Also push to history for panel rendering
+			this.history.push({
+				type: 'error',
+				errorType: 'error',
+				message,
+			})
+			this.#emitHistoryChange()
+		})
+
+		if (this.config.customTools) {
+			for (const [name, tool] of Object.entries(this.config.customTools)) {
+				if (tool === null) {
+					this.tools.delete(name)
+					continue
+				}
+				this.tools.set(name, tool)
+			}
+		}
+
+		if (!this.config.experimentalScriptExecutionTool) {
+			this.tools.delete('execute_javascript')
+		}
+	}
+
+	/** Get current agent status */
+	get status(): AgentStatus {
+		return this.#status
+	}
+
+	/** Emit statuschange event */
+	#emitStatusChange(): void {
+		this.dispatchEvent(new Event('statuschange'))
+	}
+
+	/** Emit historychange event */
+	#emitHistoryChange(): void {
+		this.dispatchEvent(new Event('historychange'))
+	}
+
+	/**
+	 * Emit activity event - for transient UI feedback
+	 * @param activity - Current agent activity
+	 */
+	emitActivity(activity: AgentActivity): void {
+		this.dispatchEvent(new CustomEvent('activity', { detail: activity }))
+	}
+
+	/** Update status and emit event */
+	#setStatus(status: AgentStatus): void {
+		if (this.#status !== status) {
+			this.#status = status
+			this.#emitStatusChange()
+		}
+	}
+
+	/**
+	 * Push a persistent observation to the history event stream.
+	 * This will be visible in <agent_history> and remain in memory across steps.
+	 */
+	pushObservation(content: string): void {
+		this.history.push({ type: 'observation', content })
+		this.#emitHistoryChange()
+	}
+
+	async execute(task: string): Promise<ExecutionResult> {
+		if (!task) throw new Error('Task is required')
+		this.task = task
+		this.taskId = uid()
+
+		// Disable ask_user tool if onAskUser is not set
+		if (!this.onAskUser) {
+			this.tools.delete('ask_user')
+		}
+
+		const onBeforeStep = this.config.onBeforeStep || (() => void 0)
+		const onAfterStep = this.config.onAfterStep || (() => void 0)
+		const onBeforeTask = this.config.onBeforeTask || (() => void 0)
+		const onAfterTask = this.config.onAfterTask || (() => void 0)
+
+		await onBeforeTask.call(this)
+
+		// Show mask
+		await this.pageController.showMask()
+
+		if (this.#abortController) {
+			this.#abortController.abort()
+			this.#abortController = new AbortController()
+		}
+
+		this.history = []
+		this.#setStatus('running')
+		this.#emitHistoryChange()
+
+		// Reset states
+		this.states = {
+			totalWaitTime: 0,
+			lastURL: '',
+		}
+
+		try {
+			let step = 0
+
+			while (true) {
+				await this.#generateObservations(step)
+
+				await onBeforeStep.call(this, step)
+
+				console.group(`step: ${step}`)
+
+				// abort
+				if (this.#abortController.signal.aborted) throw new Error('AbortError')
+
+				// Thinking
+				console.log(chalk.blue('Thinking...'))
+				this.emitActivity({ type: 'thinking' })
+
+				const result = await this.#llm.invoke(
+					[
+						{
+							role: 'system',
+							content: this.#getSystemPrompt(),
+						},
+						{
+							role: 'user',
+							content: await this.#assembleUserPrompt(),
+						},
+					],
+					{ AgentOutput: this.#packMacroTool() },
+					this.#abortController.signal,
+					{
+						toolChoiceName: 'AgentOutput',
+						normalizeResponse,
+					}
+				)
+
+				const macroResult = result.toolResult as MacroToolResult
+				const input = macroResult.input
+				const output = macroResult.output
+				const reflection: Partial<AgentReflection> = {
+					evaluation_previous_goal: input.evaluation_previous_goal,
+					memory: input.memory,
+					next_goal: input.next_goal,
+				}
+				const actionName = Object.keys(input.action)[0]
+				const action: AgentStep['action'] = {
+					name: actionName,
+					input: input.action[actionName],
+					output: output,
+				}
+
+				this.history.push({
+					type: 'step',
+					reflection,
+					action,
+					usage: result.usage,
+				} as AgentStep)
+				this.#emitHistoryChange()
+
+				console.log(chalk.green('Step finished:'), actionName)
+				console.groupEnd()
+
+				await onAfterStep.call(this, this.history)
+
+				step++
+				if (step > MAX_STEPS) {
+					this.#onDone('Step count exceeded maximum limit', false)
+					const result: ExecutionResult = {
+						success: false,
+						data: 'Step count exceeded maximum limit',
+						history: this.history,
+					}
+					await onAfterTask.call(this, result)
+					return result
+				}
+				if (actionName === 'done') {
+					const success = action.input?.success ?? false
+					const text = action.input?.text || 'no text provided'
+					console.log(chalk.green.bold('Task completed'), success, text)
+					this.#onDone(text, success)
+					const result: ExecutionResult = {
+						success,
+						data: text,
+						history: this.history,
+					}
+					await onAfterTask.call(this, result)
+					return result
+				}
+			}
+		} catch (error: unknown) {
+			console.error('Task failed', error)
+			const errorMessage = String(error)
+			this.emitActivity({ type: 'error', message: errorMessage })
+			this.#onDone(errorMessage, false)
+			const result: ExecutionResult = {
+				success: false,
+				data: errorMessage,
+				history: this.history,
+			}
+			await onAfterTask.call(this, result)
+			return result
+		}
+	}
+
+	/**
+	 * Merge all tools into a single MacroTool with the following input:
+	 * - thinking: string
+	 * - evaluation_previous_goal: string
+	 * - memory: string
+	 * - next_goal: string
+	 * - action: { toolName: toolInput }
+	 * where action must be selected from tools defined in this.tools
+	 */
+	#packMacroTool(): Tool<MacroToolInput, MacroToolResult> {
+		const tools = this.tools
+
+		const actionSchemas = Array.from(tools.entries()).map(([toolName, tool]) => {
+			return zod.object({ [toolName]: tool.inputSchema }).describe(tool.description)
+		})
+
+		const actionSchema = zod.union(
+			actionSchemas as unknown as [zod.ZodType, zod.ZodType, ...zod.ZodType[]]
+		)
+
+		const macroToolSchema = zod.object({
+			// thinking: zod.string().optional(),
+			evaluation_previous_goal: zod.string().optional(),
+			memory: zod.string().optional(),
+			next_goal: zod.string().optional(),
+			action: actionSchema,
+		})
+
+		return {
+			description: 'You MUST call this tool every step. Outputs your reflections and next action.',
+			inputSchema: macroToolSchema as zod.ZodType<MacroToolInput>,
+			execute: async (input: MacroToolInput): Promise<MacroToolResult> => {
+				// abort
+				if (this.#abortController.signal.aborted) throw new Error('AbortError')
+
+				console.log(chalk.blue.bold('MacroTool execute'), input)
+				const action = input.action
+
+				const toolName = Object.keys(action)[0]
+				const toolInput = action[toolName]
+
+				// Build reflection text, only include non-empty fields
+				const reflectionLines: string[] = []
+				if (input.evaluation_previous_goal)
+					reflectionLines.push(`✅: ${input.evaluation_previous_goal}`)
+				if (input.memory) reflectionLines.push(`💾: ${input.memory}`)
+				if (input.next_goal) reflectionLines.push(`🎯: ${input.next_goal}`)
+
+				const reflectionText = reflectionLines.length > 0 ? reflectionLines.join('\n') : ''
+
+				if (reflectionText) {
+					console.log(reflectionText)
+				}
+
+				// Find the corresponding tool
+				const tool = tools.get(toolName)
+				assert(tool, `Tool ${toolName} not found. (@note should have been caught before this!!!)`)
+
+				console.log(chalk.blue.bold(`Executing tool: ${toolName}`), toolInput)
+
+				// Emit executing activity
+				this.emitActivity({ type: 'executing', tool: toolName, input: toolInput })
+
+				const startTime = Date.now()
+
+				// Execute tool, bind `this` to PageAgent
+				const result = await tool.execute.bind(this)(toolInput)
+
+				const duration = Date.now() - startTime
+				console.log(chalk.green.bold(`Tool (${toolName}) executed for ${duration}ms`), result)
+
+				// Emit executed activity
+				this.emitActivity({
+					type: 'executed',
+					tool: toolName,
+					input: toolInput,
+					output: result,
+					duration,
+				})
+
+				// Reset wait time for non-wait tools
+				if (toolName !== 'wait') {
+					this.states.totalWaitTime = 0
+				}
+
+				// Return structured result
+				return {
+					input,
+					output: result,
+				}
+			},
+		}
+	}
+
+	/**
+	 * Get system prompt, dynamically replace language settings based on configured language
+	 */
+	#getSystemPrompt(): string {
+		let systemPrompt = SYSTEM_PROMPT
+
+		const targetLanguage = this.config.language === 'zh-CN' ? '中文' : 'English'
+		systemPrompt = systemPrompt.replace(
+			/Default working language: \*\*.*?\*\*/,
+			`Default working language: **${targetLanguage}**`
+		)
+
+		return systemPrompt
+	}
+
+	/**
+	 * Get instructions from config and format as XML block
+	 */
+	async #getInstructions(): Promise<string> {
+		const { instructions } = this.config
+		if (!instructions) return ''
+
+		const systemInstructions = instructions.system?.trim()
+		const url = await this.pageController.getCurrentUrl()
+		let pageInstructions: string | undefined
+
+		if (instructions.getPageInstructions) {
+			try {
+				pageInstructions = instructions.getPageInstructions(url)?.trim()
+			} catch (error) {
+				console.error(
+					chalk.red('[PageAgent] Failed to execute getPageInstructions callback:'),
+					error
+				)
+			}
+		}
+		if (!systemInstructions && !pageInstructions) return ''
+
+		let result = '<instructions>\n'
+
+		if (systemInstructions) {
+			result += `<system_instructions>\n${systemInstructions}\n</system_instructions>\n`
+		}
+
+		if (pageInstructions) {
+			result += `<page_instructions>\n${pageInstructions}\n</page_instructions>\n`
+		}
+
+		result += '</instructions>\n\n'
+
+		return result
+	}
+
+	/**
+	 * Generate observations before each step
+	 * - URL change detection
+	 * - Too many steps warning
+	 * @todo loop detection
+	 * @todo console error
+	 */
+	async #generateObservations(stepCount: number): Promise<void> {
+		// Detect URL change
+		const currentURL = await this.pageController.getCurrentUrl()
+		if (currentURL !== this.states.lastURL) {
+			this.pushObservation(`Page navigated to → ${currentURL}`)
+			this.states.lastURL = currentURL
+		}
+
+		// Warn about remaining steps
+		const remaining = MAX_STEPS - stepCount
+		if (remaining === 5) {
+			this.pushObservation(
+				`⚠️ Only ${remaining} steps remaining. Consider wrapping up or calling done with partial results.`
+			)
+		} else if (remaining === 2) {
+			this.pushObservation(
+				`⚠️ Critical: Only ${remaining} steps left! You must finish the task or call done immediately.`
+			)
+		}
+	}
+
+	async #assembleUserPrompt(): Promise<string> {
+		let prompt = ''
+
+		// <instructions> (optional)
+		prompt += await this.#getInstructions()
+
+		// <agent_state>
+		//  - <user_request>
+		//  - <step_info>
+		// <agent_state>
+
+		const stepCount = this.history.filter((e) => e.type === 'step').length
+
+		prompt += `<agent_state>
+			<user_request>
+			${this.task}
+			</user_request>
+			<step_info>
+			Step ${stepCount + 1} of ${MAX_STEPS} max possible steps
+			Current date and time: ${new Date().toISOString()}
+			</step_info>
+			</agent_state>
+		`
+
+		// <agent_history>
+		//  - <step_N> for steps
+		//  - <sys> for observations and system messages
+
+		prompt += '\n<agent_history>\n'
+
+		let stepIndex = 0
+		for (const event of this.history) {
+			if (event.type === 'step') {
+				stepIndex++
+				prompt += `<step_${stepIndex}>
+				Evaluation of Previous Step: ${event.reflection.evaluation_previous_goal}
+				Memory: ${event.reflection.memory}
+				Next Goal: ${event.reflection.next_goal}
+				Action Results: ${event.action.output}
+				</step_${stepIndex}>
+			`
+			} else if (event.type === 'observation') {
+				prompt += `<sys>${event.content}</sys>\n`
+			} else if (event.type === 'user_takeover') {
+				prompt += `<sys>User took over control and made changes to the page.</sys>\n`
+			} else if (event.type === 'error') {
+				// Error events are mainly for panel rendering, not included in LLM context
+				// to avoid polluting the agent's reasoning with transient errors
+			}
+		}
+
+		prompt += '</agent_history>\n\n'
+
+		// <browser_state>
+
+		prompt += await this.#getBrowserState()
+
+		return trimLines(prompt)
+	}
+
+	#onDone(text: string, success = true) {
+		this.pageController.cleanUpHighlights()
+		this.pageController.hideMask() // No await - fire and forget
+		this.#setStatus(success ? 'completed' : 'error')
+		this.#abortController.abort()
+	}
+
+	async #getBrowserState(): Promise<string> {
+		const state = await this.pageController.getBrowserState()
+
+		let content = state.content
+		if (this.config.transformPageContent) {
+			content = await this.config.transformPageContent(content)
+		}
+
+		return trimLines(`<browser_state>
+			Current Page: [${state.title}](${state.url})
+
+			${state.header}
+			${content}
+			${state.footer}
+
+			</browser_state>
+		`)
+	}
+
+	dispose(reason?: string) {
+		console.log('Disposing PageAgent...')
+		this.disposed = true
+		this.pageController.dispose()
+		this.history = []
+		this.#abortController.abort(reason ?? 'PageAgent disposed')
+
+		// Emit dispose event for UI cleanup
+		this.dispatchEvent(new Event('dispose'))
+
+		this.config.onDispose?.call(this, reason)
+	}
+}
--- a/packages/page-agent/src/config/constants.ts
+++ b/packages/page-agent/src/config/constants.ts
--- a/packages/page-agent/src/config/index.ts
+++ b/packages/page-agent/src/config/index.ts
@@ -1,42 +1,24 @@
 import type { LLMConfig } from '@page-agent/llms'
-import type { PageControllerConfig } from '@page-agent/page-controller'
-import type { SupportedLanguage } from '@page-agent/ui'
+import type { PageController, PageControllerConfig } from '@page-agent/page-controller'

-import type { ExecutionResult, HistoryEvent, PageAgent } from '../PageAgent'
+import type { PageAgentCore } from '../PageAgentCore'
 import type { PageAgentTool } from '../tools'
+import type { ExecutionResult, HistoricalEvent } from '../types'

 export type { LLMConfig }

+/** Supported UI languages */
+export type SupportedLanguage = 'en-US' | 'zh-CN'
+
 export interface AgentConfig {
 	// theme?: 'light' | 'dark'
 	language?: SupportedLanguage

-	/**
-	 * Whether to prompt for next task after task completion
-	 * @default true
-	 */
-	promptForNextTask?: boolean
-
-	/**
-	 * Enable the UI panel for visual feedback and user interaction
-	 * When disabled, the panel will not be created and all UI operations will be skipped.
-	 * Useful for automated testing or when integrating PageAgent as a library.
-	 * @default true
-	 */
-	enablePanel?: boolean
-
-	/**
-	 * Enable the ask_user tool for agent to ask questions
-	 * When disabled, the agent cannot ask user questions during execution.
-	 * @default true
-	 */
-	enableAskUser?: boolean
-
 	/**
 	 * Custom tools to extend PageAgent capabilities
 	 * @experimental
 	 * @note You can also override or remove internal tools by using the same name.
-	 * @see [tools](../tools/index.ts)
+	 * @see PageAgentTool
 	 *
 	 * @example
 	 * // override internal tool
@@ -85,17 +67,16 @@ export interface AgentConfig {
 	// @todo: use event instead of hooks
 	// @todo: remove `this` binding, pass agent as explicit parameter instead

-	onBeforeStep?: (this: PageAgent, stepCnt: number) => Promise<void> | void
-	onAfterStep?: (this: PageAgent, stepCnt: number, history: HistoryEvent[]) => Promise<void> | void
-	onBeforeTask?: (this: PageAgent) => Promise<void> | void
-	onAfterTask?: (this: PageAgent, result: ExecutionResult) => Promise<void> | void
+	onBeforeStep?: (this: PageAgentCore, stepCnt: number) => Promise<void> | void
+	onAfterStep?: (this: PageAgentCore, history: HistoricalEvent[]) => Promise<void> | void
+	onBeforeTask?: (this: PageAgentCore) => Promise<void> | void
+	onAfterTask?: (this: PageAgentCore, result: ExecutionResult) => Promise<void> | void

 	/**
 	 * @note this hook can block the disposal process
-	 * @note when dispose caused by page unload, reason will be 'PAGE_UNLOADING'. this method CANNOT block unloading. async operations may be cut.
 	 * @todo remove `this` binding, pass agent as explicit parameter instead
 	 */
-	onDispose?: (this: PageAgent, reason?: string) => void
+	onDispose?: (this: PageAgentCore, reason?: string) => void

 	// page behavior hooks

--- a/packages/core/src/env.d.ts
+++ b/packages/core/src/env.d.ts
@@ -0,0 +1,6 @@
+/// <reference types="vite/client" />
+
+declare module '*.md?raw' {
+	const content: string
+	export default content
+}
--- a/packages/page-agent/src/prompts/.prettierignore
+++ b/packages/page-agent/src/prompts/.prettierignore
--- a/packages/page-agent/src/prompts/system_prompt.md
+++ b/packages/page-agent/src/prompts/system_prompt.md
--- a/packages/page-agent/src/tools/index.ts
+++ b/packages/page-agent/src/tools/index.ts
@@ -4,7 +4,7 @@
 */
 import zod, { type z } from 'zod'

-import type { PageAgent } from '../PageAgent'
+import type { PageAgentCore } from '../PageAgentCore'
 import { waitFor } from '../utils'

 /**
@@ -14,7 +14,7 @@ export interface PageAgentTool<TParams = any> {
 	// name: string
 	description: string
 	inputSchema: z.ZodType<TParams>
-	execute: (this: PageAgent, args: TParams) => Promise<string>
+	execute: (this: PageAgentCore, args: TParams) => Promise<string>
 }

 export function tool<TParams>(options: PageAgentTool<TParams>): PageAgentTool<TParams> {
@@ -36,7 +36,7 @@ tools.set(
 			text: zod.string(),
 			success: zod.boolean().default(true),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			// @note main loop will handle this one
 			// this.onDone(input.text, input.success)
 			return Promise.resolve('Task completed')
@@ -52,7 +52,7 @@ tools.set(
 		inputSchema: zod.object({
 			seconds: zod.number().min(1).max(10).default(1),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			const lastTimeUpdate = await this.pageController.getLastUpdateTime()
 			const actualWaitTime = Math.max(0, input.seconds - (Date.now() - lastTimeUpdate) / 1000)
 			console.log(`actualWaitTime: ${actualWaitTime} seconds`)
@@ -79,12 +79,12 @@ tools.set(
 		inputSchema: zod.object({
 			question: zod.string(),
 		}),
-		execute: async function (this: PageAgent, input) {
-			if (!this.panel) {
-				throw new Error('ask_user tool requires panel to be enabled')
+		execute: async function (this: PageAgentCore, input) {
+			if (!this.onAskUser) {
+				throw new Error('ask_user tool requires onAskUser callback to be set')
 			}
-			const answer = await this.panel.askUser(input.question)
-			return `✅ Received user answer: ${answer}`
+			const answer = await this.onAskUser(input.question)
+			return `User answered: ${answer}`
 		},
 	})
 )
@@ -96,7 +96,7 @@ tools.set(
 		inputSchema: zod.object({
 			index: zod.int().min(0),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			const result = await this.pageController.clickElement(input.index)
 			return result.message
 		},
@@ -111,7 +111,7 @@ tools.set(
 			index: zod.int().min(0),
 			text: zod.string(),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			const result = await this.pageController.inputText(input.index, input.text)
 			return result.message
 		},
@@ -127,7 +127,7 @@ tools.set(
 			index: zod.int().min(0),
 			text: zod.string(),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			const result = await this.pageController.selectOption(input.index, input.text)
 			return result.message
 		},
@@ -148,7 +148,7 @@ tools.set(
 			pixels: zod.number().int().min(0).optional(),
 			index: zod.number().int().min(0).optional(),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			const result = await this.pageController.scroll({
 				...input,
 				numPages: input.num_pages,
@@ -168,7 +168,7 @@ tools.set(
 			pixels: zod.number().int().min(0),
 			index: zod.number().int().min(0).optional(),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			const result = await this.pageController.scrollHorizontally(input)
 			return result.message
 		},
@@ -183,7 +183,7 @@ tools.set(
 		inputSchema: zod.object({
 			script: zod.string(),
 		}),
-		execute: async function (this: PageAgent, input) {
+		execute: async function (this: PageAgentCore, input) {
 			const result = await this.pageController.executeJavascript(input.script)
 			return result.message
 		},
--- a/packages/core/src/types.ts
+++ b/packages/core/src/types.ts
@@ -0,0 +1,109 @@
+/**
+ * Agent reflection 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
+}
+
+/**
+ * A single agent step with reflection and action
+ */
+export interface AgentStep {
+	type: 'step'
+	reflection: Partial<AgentReflection>
+	action: {
+		name: string
+		input: any
+		output: string
+	}
+	usage: {
+		promptTokens: number
+		completionTokens: number
+		totalTokens: number
+		cachedTokens?: number
+		reasoningTokens?: number
+	}
+}
+
+/**
+ * Persistent observation event (stays in memory)
+ */
+export interface ObservationEvent {
+	type: 'observation'
+	content: string
+}
+
+/**
+ * User takeover event
+ */
+export interface UserTakeoverEvent {
+	type: 'user_takeover'
+}
+
+/**
+ * Error event (retry or error from LLM)
+ */
+export interface ErrorEvent {
+	type: 'error'
+	errorType: 'retry' | 'error'
+	message: string
+	attempt?: number
+	maxAttempts?: number
+}
+
+/**
+ * Union type for all history events
+ */
+export type HistoricalEvent = AgentStep | ObservationEvent | UserTakeoverEvent | ErrorEvent
+
+/**
+ * Agent execution status
+ */
+export type AgentStatus = 'idle' | 'running' | 'completed' | 'error'
+
+/**
+ * Agent activity - transient state for immediate UI feedback.
+ *
+ * Unlike historical events (which are persisted), activities are ephemeral
+ * and represent "what the agent is doing right now". UI components should
+ * listen to 'activity' events to show real-time feedback.
+ *
+ * Note: There is no 'idle' activity - absence of activity events means idle.
+ */
+export type AgentActivity =
+	| { type: 'thinking' }
+	| { type: 'executing'; tool: string; input: unknown }
+	| { type: 'executed'; tool: string; input: unknown; output: string; duration: number }
+	| { type: 'retrying'; attempt: number; maxAttempts: number }
+	| { type: 'error'; message: string }
+
+export interface ExecutionResult {
+	success: boolean
+	data: string
+	history: HistoricalEvent[]
+}
--- a/packages/page-agent/src/utils/assert.ts
+++ b/packages/page-agent/src/utils/assert.ts
--- a/packages/page-agent/src/utils/autoFixer.ts
+++ b/packages/page-agent/src/utils/autoFixer.ts
--- a/packages/page-agent/src/utils/index.ts
+++ b/packages/page-agent/src/utils/index.ts
@@ -68,12 +68,13 @@ export function randomID(existingIDs?: string[]): string {
 }

 //
+const _global = globalThis as any

-if (!window.__PAGE_AGENT_IDS__) {
-	window.__PAGE_AGENT_IDS__ = []
+if (!_global.__PAGE_AGENT_IDS__) {
+	_global.__PAGE_AGENT_IDS__ = []
 }

-const ids = window.__PAGE_AGENT_IDS__
+const ids = _global.__PAGE_AGENT_IDS__

 /**
 * Generate a random ID.
--- a/packages/core/tsconfig.dts.json
+++ b/packages/core/tsconfig.dts.json
@@ -0,0 +1,9 @@
+{
+    "extends": "./tsconfig.json",
+    "compilerOptions": {
+        // @workaround DTS bug
+        // dts do not work with monorepo path mapping
+        // disable path mapping for it
+        "paths": {}
+    }
+}
--- a/packages/core/tsconfig.json
+++ b/packages/core/tsconfig.json
@@ -0,0 +1,22 @@
+{
+    "extends": "../../tsconfig.base.json",
+    "compilerOptions": {
+        "tsBuildInfoFile": "./node_modules/.tmp/tsconfig.tsbuildinfo",
+        "noEmit": false,
+        "allowImportingTsExtensions": false,
+        "baseUrl": ".",
+        "outDir": "dist",
+        "paths": {
+            //
+            "@page-agent/llms": ["../llms/src/index.ts"],
+            "@page-agent/page-controller": ["../page-controller/src/PageController.ts"]
+        }
+    },
+    "include": ["**/*.ts"],
+    "exclude": ["dist", "node_modules"],
+    "references": [
+        //
+        { "path": "../llms" },
+        { "path": "../page-controller" }
+    ]
+}
--- a/packages/core/vite.config.js
+++ b/packages/core/vite.config.js
@@ -0,0 +1,44 @@
+// @ts-check
+import { dirname, resolve } from 'path'
+import dts from 'unplugin-dts/vite'
+import { fileURLToPath } from 'url'
+import { defineConfig } from 'vite'
+import cssInjectedByJsPlugin from 'vite-plugin-css-injected-by-js'
+
+const __dirname = dirname(fileURLToPath(import.meta.url))
+
+// ES Module for NPM Package
+export default defineConfig({
+	clearScreen: false,
+	plugins: [
+		dts({ tsconfigPath: './tsconfig.dts.json', bundleTypes: true }),
+		cssInjectedByJsPlugin({ relativeCSSInjection: true }),
+	],
+	publicDir: false,
+	esbuild: {
+		keepNames: true,
+	},
+	build: {
+		lib: {
+			entry: resolve(__dirname, 'src/PageAgentCore.ts'),
+			name: 'PageAgentCore',
+			fileName: 'page-agent-core',
+			formats: ['es'],
+		},
+		outDir: resolve(__dirname, 'dist', 'esm'),
+		rollupOptions: {
+			external: [
+				'chalk',
+				'zod',
+				// all the internal packages
+				/^@page-agent\//,
+			],
+		},
+		minify: false,
+		sourcemap: true,
+		cssCodeSplit: true,
+	},
+	define: {
+		'process.env.NODE_ENV': '"production"',
+	},
+})
--- a/packages/llms/src/index.ts
+++ b/packages/llms/src/index.ts
@@ -56,9 +56,9 @@ export class LLM extends EventTarget {
 			// retry settings
 			{
 				maxRetries: this.config.maxRetries,
-				onRetry: (current: number) => {
+				onRetry: (attempt: number) => {
 					this.dispatchEvent(
-						new CustomEvent('retry', { detail: { current, max: this.config.maxRetries } })
+						new CustomEvent('retry', { detail: { attempt, maxAttempts: this.config.maxRetries } })
 					)
 				},
 				onError: (error: Error) => {
@@ -73,15 +73,15 @@ async function withRetry<T>(
 	fn: () => Promise<T>,
 	settings: {
 		maxRetries: number
-		onRetry: (retries: number) => void
+		onRetry: (attempt: number) => void
 		onError: (error: Error) => void
 	}
 ): Promise<T> {
-	let retries = 0
+	let attempt = 0
 	let lastError: Error | null = null
-	while (retries <= settings.maxRetries) {
-		if (retries > 0) {
-			settings.onRetry(retries)
+	while (attempt <= settings.maxRetries) {
+		if (attempt > 0) {
+			settings.onRetry(attempt)
 			await new Promise((resolve) => setTimeout(resolve, 100))
 		}

@@ -98,7 +98,7 @@ async function withRetry<T>(
 			if (error instanceof InvokeError && !error.retryable) throw error

 			lastError = error as Error
-			retries++
+			attempt++

 			await new Promise((resolve) => setTimeout(resolve, 100))
 		}
--- a/packages/page-agent/package.json
+++ b/packages/page-agent/package.json
@@ -47,6 +47,7 @@
        "zod": "^4.3.5",
        "@page-agent/llms": "0.2.5",
        "@page-agent/page-controller": "0.2.5",
+        "@page-agent/core": "0.2.5",
        "@page-agent/ui": "0.2.5"
    }
 }
--- a/packages/page-agent/src/PageAgent.ts
+++ b/packages/page-agent/src/PageAgent.ts
@@ -2,628 +2,25 @@
 * Copyright (C) 2025 Alibaba Group Holding Limited
 * All rights reserved.
 */
-import { LLM, type Tool } from '@page-agent/llms'
+import { type PageAgentConfig, PageAgentCore } from '@page-agent/core'
 import { PageController } from '@page-agent/page-controller'
 import { Panel } from '@page-agent/ui'
-import chalk from 'chalk'
-import zod from 'zod'
-
-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 { normalizeResponse, trimLines, uid } from './utils'
-import { assert } from './utils/assert'
-
-/**
- * Agent reflection 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'

-/**
- * A single agent step with reflection and action
- */
-export interface AgentStep {
-	type: 'step'
-	reflection: Partial<AgentReflection>
-	action: {
-		name: string
-		input: any
-		output: string
-	}
-	usage: {
-		promptTokens: number
-		completionTokens: number
-		totalTokens: number
-		cachedTokens?: number
-		reasoningTokens?: number
-	}
-}
-
-/**
- * Persistent observation event (stays in memory)
- */
-export interface ObservationEvent {
-	type: 'observation'
-	content: string
-}
-
-/**
- * User takeover event
- */
-export interface UserTakeoverEvent {
-	type: 'user_takeover'
-}
-
-/**
- * Union type for all history events
- */
-export type HistoryEvent = AgentStep | ObservationEvent | UserTakeoverEvent
-
-export interface ExecutionResult {
-	success: boolean
-	data: string
-	history: HistoryEvent[]
-}
-
-export class PageAgent extends EventTarget {
-	config: PageAgentConfig
-	id = uid()
-	panel: Panel | null = null
-	tools: typeof tools
-	disposed = false
-	task = ''
-	taskId = ''
-
-	#llm: LLM
-	#abortController = new AbortController()
-	#llmRetryListener: ((e: Event) => void) | null = null
-	#llmErrorListener: ((e: Event) => void) | null = null
-	#beforeUnloadListener: ((e: Event) => void) | null = null
-
-	/** PageController for DOM operations */
-	pageController: PageController
-
-	/** Runtime states for tracking across steps */
-	states = {
-		/** Accumulated wait time in seconds, used by wait tool */
-		totalWaitTime: 0,
-		/** Last known URL for detecting navigation */
-		lastURL: '',
-	}
-
-	/** History events */
-	history: HistoryEvent[] = []
+export class PageAgent extends PageAgentCore {
+	panel: Panel

 	constructor(config: PageAgentConfig) {
-		super()
-
-		this.config = config
-		this.#llm = new LLM(this.config)
-
-		// Conditionally initialize Panel
-		if (this.config.enablePanel !== false) {
-			this.panel = new Panel({
-				language: this.config.language,
-				onExecuteTask: (task) => this.execute(task),
-				onStop: () => this.dispose(),
-				promptForNextTask: this.config.promptForNextTask,
-			})
-		}
-
-		this.tools = new Map(tools)
-
-		// Initialize PageController with config (mask enabled by default)
-		this.pageController = new PageController({
-			...this.config,
-			enableMask: this.config.enableMask ?? true,
+		const pageController = new PageController({
+			...config,
+			enableMask: config.enableMask ?? true,
 		})

-		// Listen to LLM events
-		this.#llmRetryListener = (e) => {
-			const { current, max } = (e as CustomEvent).detail
-			this.panel?.update({ type: 'retry', current, max })
-		}
-		this.#llmErrorListener = (e) => {
-			const { error } = (e as CustomEvent).detail
-			this.panel?.update({ type: 'error', message: `step failed: ${error.message}` })
-		}
-		this.#llm.addEventListener('retry', this.#llmRetryListener)
-		this.#llm.addEventListener('error', this.#llmErrorListener)
+		super({ ...config, pageController })

-		if (this.config.customTools) {
-			for (const [name, tool] of Object.entries(this.config.customTools)) {
-				if (tool === null) {
-					this.tools.delete(name)
-					continue
-				}
-				this.tools.set(name, tool)
-			}
-		}
-
-		if (!this.config.experimentalScriptExecutionTool) {
-			this.tools.delete('execute_javascript')
-		}
-
-		// Disable ask_user tool if enableAskUser is false or if panel is disabled
-		if (this.config.enableAskUser === false || this.config.enablePanel === false) {
-			this.tools.delete('ask_user')
-		}
-
-		this.#beforeUnloadListener = (e) => {
-			if (!this.disposed) this.dispose('PAGE_UNLOADING')
-		}
-		window.addEventListener('beforeunload', this.#beforeUnloadListener)
-	}
-
-	/**
-	 * Push a persistent observation to the history event stream.
-	 * This will be visible in <agent_history> and remain in memory across steps.
-	 */
-	pushObservation(content: string): void {
-		this.history.push({ type: 'observation', content })
-		this.panel?.update({ type: 'observation', content })
-	}
-
-	async execute(task: string): Promise<ExecutionResult> {
-		if (!task) throw new Error('Task is required')
-		this.task = task
-		this.taskId = uid()
-
-		const onBeforeStep = this.config.onBeforeStep || (() => void 0)
-		const onAfterStep = this.config.onAfterStep || (() => void 0)
-		const onBeforeTask = this.config.onBeforeTask || (() => void 0)
-		const onAfterTask = this.config.onAfterTask || (() => void 0)
-
-		await onBeforeTask.call(this)
-
-		// Show mask and panel
-		this.pageController.showMask()
-
-		this.panel?.show()
-		this.panel?.reset()
-
-		this.panel?.update({ type: 'input', task: this.task })
-
-		if (this.#abortController) {
-			this.#abortController.abort()
-			this.#abortController = new AbortController()
-		}
-
-		this.history = []
-
-		// Reset states
-		this.states = {
-			totalWaitTime: 0,
-			lastURL: '',
-		}
-
-		try {
-			let step = 0
-
-			while (true) {
-				await this.#generateObservations(step)
-
-				await onBeforeStep.call(this, step)
-
-				console.group(`step: ${step}`)
-
-				// abort
-				if (this.#abortController.signal.aborted) throw new Error('AbortError')
-
-				// Update status to thinking
-				console.log(chalk.blue('Thinking...'))
-				this.panel?.update({ type: 'thinking' })
-
-				const result = await this.#llm.invoke(
-					[
-						{
-							role: 'system',
-							content: this.#getSystemPrompt(),
-						},
-						{
-							role: 'user',
-							content: await this.#assembleUserPrompt(),
-						},
-					],
-					{ AgentOutput: this.#packMacroTool() },
-					this.#abortController.signal,
-					{
-						toolChoiceName: 'AgentOutput',
-						normalizeResponse,
-					}
-				)
-
-				const macroResult = result.toolResult as MacroToolResult
-				const input = macroResult.input
-				const output = macroResult.output
-				const reflection: Partial<AgentReflection> = {
-					evaluation_previous_goal: input.evaluation_previous_goal,
-					memory: input.memory,
-					next_goal: input.next_goal,
-				}
-				const actionName = Object.keys(input.action)[0]
-				const action: AgentStep['action'] = {
-					name: actionName,
-					input: input.action[actionName],
-					output: output,
-				}
-
-				this.history.push({
-					type: 'step',
-					reflection,
-					action,
-					usage: result.usage,
-				} as AgentStep)
-
-				console.log(chalk.green('Step finished:'), actionName)
-				console.groupEnd()
-
-				await onAfterStep.call(this, step, this.history)
-
-				step++
-				if (step > MAX_STEPS) {
-					this.#onDone('Step count exceeded maximum limit', false)
-					const result: ExecutionResult = {
-						success: false,
-						data: 'Step count exceeded maximum limit',
-						history: this.history,
-					}
-					await onAfterTask.call(this, result)
-					return result
-				}
-				if (actionName === 'done') {
-					const success = action.input?.success ?? false
-					const text = action.input?.text || 'no text provided'
-					console.log(chalk.green.bold('Task completed'), success, text)
-					this.#onDone(text, success)
-					const result: ExecutionResult = {
-						success,
-						data: text,
-						history: this.history,
-					}
-					await onAfterTask.call(this, result)
-					return result
-				}
-			}
-		} catch (error: unknown) {
-			console.error('Task failed', error)
-			this.#onDone(String(error), false)
-			const result: ExecutionResult = {
-				success: false,
-				data: String(error),
-				history: this.history,
-			}
-			await onAfterTask.call(this, result)
-			return result
-		}
-	}
-
-	/**
-	 * Merge all tools into a single MacroTool with the following input:
-	 * - thinking: string
-	 * - evaluation_previous_goal: string
-	 * - memory: string
-	 * - next_goal: string
-	 * - action: { toolName: toolInput }
-	 * where action must be selected from tools defined in this.tools
-	 */
-	#packMacroTool(): Tool<MacroToolInput, MacroToolResult> {
-		const tools = this.tools
-
-		const actionSchemas = Array.from(tools.entries()).map(([toolName, tool]) => {
-			return zod.object({ [toolName]: tool.inputSchema }).describe(tool.description)
+		this.panel = new Panel(this, {
+			language: config.language,
 		})
-
-		const actionSchema = zod.union(
-			actionSchemas as unknown as [zod.ZodType, zod.ZodType, ...zod.ZodType[]]
-		)
-
-		const macroToolSchema = zod.object({
-			// thinking: zod.string().optional(),
-			evaluation_previous_goal: zod.string().optional(),
-			memory: zod.string().optional(),
-			next_goal: zod.string().optional(),
-			action: actionSchema,
-		})
-
-		return {
-			description: 'You MUST call this tool every step. Outputs your reflections and next action.',
-			inputSchema: macroToolSchema as zod.ZodType<MacroToolInput>,
-			execute: async (input: MacroToolInput): Promise<MacroToolResult> => {
-				// abort
-				if (this.#abortController.signal.aborted) throw new Error('AbortError')
-
-				console.log(chalk.blue.bold('MacroTool execute'), input)
-				const action = input.action
-
-				const toolName = Object.keys(action)[0]
-				const toolInput = action[toolName]
-
-				// Build reflection text, only include non-empty fields
-				const reflectionLines: string[] = []
-				if (input.evaluation_previous_goal)
-					reflectionLines.push(`✅: ${input.evaluation_previous_goal}`)
-				if (input.memory) reflectionLines.push(`💾: ${input.memory}`)
-				if (input.next_goal) reflectionLines.push(`🎯: ${input.next_goal}`)
-
-				const reflectionText = reflectionLines.length > 0 ? reflectionLines.join('\n') : ''
-
-				if (reflectionText) {
-					console.log(reflectionText)
-					this.panel?.update({ type: 'thinking', text: reflectionText })
-				}
-
-				// Find the corresponding tool
-				const tool = tools.get(toolName)
-				assert(tool, `Tool ${toolName} not found. (@note should have been caught before this!!!)`)
-
-				console.log(chalk.blue.bold(`Executing tool: ${toolName}`), toolInput)
-				this.panel?.update({ type: 'toolExecuting', toolName, args: toolInput })
-
-				const startTime = Date.now()
-
-				// Execute tool, bind `this` to PageAgent
-				const result = await tool.execute.bind(this)(toolInput)
-
-				const duration = Date.now() - startTime
-				console.log(chalk.green.bold(`Tool (${toolName}) executed for ${duration}ms`), result)
-
-				// Reset wait time for non-wait tools
-				if (toolName !== 'wait') {
-					this.states.totalWaitTime = 0
-				}
-
-				// Briefly display execution result
-				this.panel?.update({
-					type: 'toolCompleted',
-					toolName,
-					args: toolInput,
-					result,
-					duration,
-				})
-
-				// Wait a moment to let user see the result
-				await new Promise((resolve) => setTimeout(resolve, 100))
-
-				// Return structured result
-				return {
-					input,
-					output: result,
-				}
-			},
-		}
-	}
-
-	/**
-	 * Get system prompt, dynamically replace language settings based on configured language
-	 */
-	#getSystemPrompt(): string {
-		let systemPrompt = SYSTEM_PROMPT
-
-		const targetLanguage = this.config.language === 'zh-CN' ? '中文' : 'English'
-		systemPrompt = systemPrompt.replace(
-			/Default working language: \*\*.*?\*\*/,
-			`Default working language: **${targetLanguage}**`
-		)
-
-		return systemPrompt
-	}
-
-	/**
-	 * Get instructions from config and format as XML block
-	 */
-	async #getInstructions(): Promise<string> {
-		const { instructions } = this.config
-		if (!instructions) return ''
-
-		const systemInstructions = instructions.system?.trim()
-		const url = await this.pageController.getCurrentUrl()
-		let pageInstructions: string | undefined
-
-		if (instructions.getPageInstructions) {
-			try {
-				pageInstructions = instructions.getPageInstructions(url)?.trim()
-			} catch (error) {
-				console.error(
-					chalk.red('[PageAgent] Failed to execute getPageInstructions callback:'),
-					error
-				)
-			}
-		}
-		if (!systemInstructions && !pageInstructions) return ''
-
-		let result = '<instructions>\n'
-
-		if (systemInstructions) {
-			result += `<system_instructions>\n${systemInstructions}\n</system_instructions>\n`
-		}
-
-		if (pageInstructions) {
-			result += `<page_instructions>\n${pageInstructions}\n</page_instructions>\n`
-		}
-
-		result += '</instructions>\n\n'
-
-		return result
-	}
-
-	/**
-	 * Generate observations before each step
-	 * - URL change detection
-	 * - Too many steps warning
-	 * @todo loop detection
-	 * @todo console error
-	 */
-	async #generateObservations(stepCount: number): Promise<void> {
-		// Detect URL change
-		const currentURL = await this.pageController.getCurrentUrl()
-		if (currentURL !== this.states.lastURL) {
-			this.pushObservation(`Page navigated to → ${currentURL}`)
-			this.states.lastURL = currentURL
-		}
-
-		// Warn about remaining steps
-		const remaining = MAX_STEPS - stepCount
-		if (remaining === 5) {
-			this.pushObservation(
-				`⚠️ Only ${remaining} steps remaining. Consider wrapping up or calling done with partial results.`
-			)
-		} else if (remaining === 2) {
-			this.pushObservation(
-				`⚠️ Critical: Only ${remaining} steps left! You must finish the task or call done immediately.`
-			)
-		}
-	}
-
-	async #assembleUserPrompt(): Promise<string> {
-		let prompt = ''
-
-		// <instructions> (optional)
-		prompt += await this.#getInstructions()
-
-		// <agent_state>
-		//  - <user_request>
-		//  - <step_info>
-		// <agent_state>
-
-		const stepCount = this.history.filter((e) => e.type === 'step').length
-
-		prompt += `<agent_state>
-			<user_request>
-			${this.task}
-			</user_request>
-			<step_info>
-			Step ${stepCount + 1} of ${MAX_STEPS} max possible steps
-			Current date and time: ${new Date().toISOString()}
-			</step_info>
-			</agent_state>
-		`
-
-		// <agent_history>
-		//  - <step_N> for steps
-		//  - <sys> for observations and system messages
-
-		prompt += '\n<agent_history>\n'
-
-		let stepIndex = 0
-		for (const event of this.history) {
-			if (event.type === 'step') {
-				stepIndex++
-				prompt += `<step_${stepIndex}>
-				Evaluation of Previous Step: ${event.reflection.evaluation_previous_goal}
-				Memory: ${event.reflection.memory}
-				Next Goal: ${event.reflection.next_goal}
-				Action Results: ${event.action.output}
-				</step_${stepIndex}>
-			`
-			} else if (event.type === 'observation') {
-				prompt += `<sys>${event.content}</sys>\n`
-			} else if (event.type === 'user_takeover') {
-				prompt += `<sys>User took over control and made changes to the page.</sys>\n`
-			}
-		}
-
-		prompt += '</agent_history>\n\n'
-
-		// <browser_state>
-
-		prompt += await this.#getBrowserState()
-
-		return trimLines(prompt)
-	}
-
-	#onDone(text: string, success = true) {
-		this.pageController.cleanUpHighlights()
-
-		// Update panel status
-		if (success) {
-			this.panel?.update({ type: 'output', text })
-		} else {
-			this.panel?.update({ type: 'error', message: text })
-		}
-
-		// Task completed
-		this.panel?.update({ type: 'completed' })
-
-		this.pageController.hideMask()
-
-		this.#abortController.abort()
-	}
-
-	async #getBrowserState(): Promise<string> {
-		const state = await this.pageController.getBrowserState()
-
-		let content = state.content
-		if (this.config.transformPageContent) {
-			content = await this.config.transformPageContent(content)
-		}
-
-		return trimLines(`<browser_state>
-			Current Page: [${state.title}](${state.url})
-
-			${state.header}
-			${content}
-			${state.footer}
-
-			</browser_state>
-		`)
-	}
-
-	dispose(reason?: string) {
-		console.log('Disposing PageAgent...')
-		this.disposed = true
-		this.pageController.dispose()
-		this.panel?.dispose()
-		this.history = []
-		this.#abortController.abort(reason ?? 'PageAgent disposed')
-
-		// Clean up LLM event listeners
-		if (this.#llmRetryListener) {
-			this.#llm.removeEventListener('retry', this.#llmRetryListener)
-			this.#llmRetryListener = null
-		}
-		if (this.#llmErrorListener) {
-			this.#llm.removeEventListener('error', this.#llmErrorListener)
-			this.#llmErrorListener = null
-		}
-
-		// Clean up window event listeners
-		if (this.#beforeUnloadListener) {
-			window.removeEventListener('beforeunload', this.#beforeUnloadListener)
-			this.#beforeUnloadListener = null
-		}
-
-		this.config.onDispose?.call(this, reason)
 	}
 }
--- a/packages/page-agent/src/env.d.ts
+++ b/packages/page-agent/src/env.d.ts
@@ -1,20 +1,9 @@
 /// <reference types="vite/client" />
 import type { PageAgent } from './PageAgent'

-declare module '*.module.css' {
-	const classes: Record<string, string>
-	export default classes
-}
-
-declare module '*.md?raw' {
-	const content: string
-	export default content
-}
-
 declare global {
 	interface Window {
 		pageAgent?: PageAgent
 		PageAgent: typeof PageAgent
-		__PAGE_AGENT_IDS__: string[]
 	}
 }
--- a/packages/page-agent/src/iife.ts
+++ b/packages/page-agent/src/iife.ts
@@ -1,6 +1,8 @@
 /**
 * Auto-run entry for page-agent.js. Insert this script into your page to get page-agent functionality.
 */
+import { Panel } from '@page-agent/ui'
+
 import { PageAgent, type PageAgentConfig } from './PageAgent'

 // Clean up existing instances to prevent multiple injections from bookmarklet
@@ -24,6 +26,8 @@ const DEMO_API_KEY = 'PAGE-AGENT-FREE-TESTING-RANDOM'
 // @todo give a switch to disable auto-init
 setTimeout(() => {
 	const currentScript = document.currentScript as HTMLScriptElement | null
+	let config: PageAgentConfig
+
 	if (currentScript) {
 		console.log('🚀 page-agent.js detected current script:', currentScript.src)
 		const url = new URL(currentScript.src)
@@ -31,23 +35,18 @@ setTimeout(() => {
 		const baseURL = url.searchParams.get('baseURL') || DEMO_BASE_URL
 		const apiKey = url.searchParams.get('apiKey') || DEMO_API_KEY
 		const language = (url.searchParams.get('lang') as 'zh-CN' | 'en-US') || 'zh-CN'
-		const config: PageAgentConfig = { model, baseURL, apiKey, language }
-		window.pageAgent = new PageAgent(config)
+		config = { model, baseURL, apiKey, language }
 	} else {
 		console.log('🚀 page-agent.js no current script detected, using default demo config')
-		const config: PageAgentConfig = {
-			// model: DEMO_MODEL,
-			// baseURL: DEMO_BASE_URL,
-			// apiKey: DEMO_API_KEY,
-
+		config = {
 			model: import.meta.env.LLM_MODEL_NAME ? import.meta.env.LLM_MODEL_NAME : DEMO_MODEL,
 			baseURL: import.meta.env.LLM_BASE_URL ? import.meta.env.LLM_BASE_URL : DEMO_BASE_URL,
 			apiKey: import.meta.env.LLM_API_KEY ? import.meta.env.LLM_API_KEY : DEMO_API_KEY,
 		}
-		window.pageAgent = new PageAgent(config)
 	}

-	console.log('🚀 page-agent.js initialized with config:', window.pageAgent.config)
+	// Create agent
+	window.pageAgent = new PageAgent(config)

-	window.pageAgent.panel!.show() // Show panel
+	console.log('🚀 page-agent.js initialized with config:', window.pageAgent.config)
 })
--- a/packages/page-agent/src/utils/bus.ts
+++ b/packages/page-agent/src/utils/bus.ts
@@ -1,98 +0,0 @@
-/**
- * Event mapping definitions
- * @note Event bus callbacks must be repeatable without errors
- */
-export interface PageAgentEventMap {
-	// PageAgent status events
-	// 'agent:execute': { params: { task: string } }
-	// 'agent:done': { params: { text: string; success: boolean } }
-	// 'agent:disposed': { params: undefined }
-	// 'agent:error': { params: { error: string | Error } }
-
-	// Task status change events
-	'task:start': { params: { task: string } }
-	// 'task:complete': { params: { text: string; success: boolean } }
-	// 'task:error': { params: { error: string | Error } }
-
-	// Index signature for dynamic event names
-	// [key: string]: { params: any }
-}
-
-/**
- * Event handler type definitions
- */
-export type EventHandler<T extends keyof PageAgentEventMap> =
-	PageAgentEventMap[T]['params'] extends undefined
-		? () => void
-		: (params: PageAgentEventMap[T]['params']) => void
-
-/**
- * Async event handler type definitions
- */
-export type AsyncEventHandler<T extends keyof PageAgentEventMap> =
-	PageAgentEventMap[T]['params'] extends undefined
-		? () => Promise<void>
-		: (params: PageAgentEventMap[T]['params']) => Promise<void>
-
-/**
- * Type-safe event bus
- * @note Mainly used to decouple logic and UI
- * @note All modules of a PageAgent instance share the same EventBus instance for communication
- * @note Use with caution if delivery guarantee is needed for logic communication
- * @note `on` `once` `emit` methods handle built-in events with type protection, use `addEventListener` for other events
- */
-class EventBus extends EventTarget {
-	/**
-	 * Listen to built-in events
-	 */
-	on<T extends keyof PageAgentEventMap>(event: T, handler: EventHandler<T>): void {
-		const wrappedHandler = (e: Event) => {
-			const customEvent = e as CustomEvent
-			const params = customEvent.detail?.[0]
-			return handler(params)
-		}
-		this.addEventListener(event, wrappedHandler)
-	}
-
-	/**
-	 * Listen to built-in events (one-time)
-	 */
-	once<T extends keyof PageAgentEventMap>(event: T, handler: EventHandler<T>): void {
-		const wrappedHandler = (e: Event) => {
-			const customEvent = e as CustomEvent
-			const params = customEvent.detail?.[0]
-			return handler(params)
-		}
-		this.addEventListener(event, wrappedHandler, { once: true })
-	}
-
-	/**
-	 * Emit built-in events
-	 */
-	emit<T extends keyof PageAgentEventMap>(
-		event: T,
-		...args: PageAgentEventMap[T]['params'] extends undefined
-			? []
-			: [PageAgentEventMap[T]['params']]
-	): void {
-		const customEvent = new CustomEvent(event, { detail: args })
-		this.dispatchEvent(customEvent)
-		return
-	}
-}
-
-const buses = new Map<string, EventBus>()
-
-/**
- * Get the event bus for a given channel
- */
-export function getEventBus(channel: string) {
-	if (buses.has(channel)) {
-		return buses.get(channel)!
-	}
-	const bus = new EventBus()
-	buses.set(channel, bus)
-	return bus
-}
-
-export type { EventBus }
--- a/packages/page-agent/tsconfig.json
+++ b/packages/page-agent/tsconfig.json
@@ -10,6 +10,7 @@
            //
            "@page-agent/llms": ["../llms/src/index.ts"],
            "@page-agent/page-controller": ["../page-controller/src/PageController.ts"],
+            "@page-agent/core": ["../core/src/PageAgentCore.ts"],
            "@page-agent/ui": ["../ui/src/index.ts"]
        }
    },
@@ -19,6 +20,7 @@
        //
        { "path": "../llms" },
        { "path": "../page-controller" },
+        { "path": "../core" },
        { "path": "../ui" }
    ]
 }
--- a/packages/page-agent/vite.iife.config.js
+++ b/packages/page-agent/vite.iife.config.js
@@ -24,6 +24,7 @@ export default defineConfig(({ mode }) => ({
 		alias: {
 			'@page-agent/page-controller': resolve(__dirname, '../page-controller/src/PageController.ts'),
 			'@page-agent/llms': resolve(__dirname, '../llms/src/index.ts'),
+			'@page-agent/core': resolve(__dirname, '../core/src/PageAgentCore.ts'),
 			'@page-agent/ui': resolve(__dirname, '../ui/src/index.ts'),
 		},
 	},
--- a/packages/page-controller/src/PageController.ts
+++ b/packages/page-controller/src/PageController.ts
@@ -18,11 +18,8 @@ import { VIEWPORT_EXPANSION } from './constants'
 import * as dom from './dom'
 import type { FlatDomTree, InteractiveElementDomNode } from './dom/dom_tree/type'
 import { getPageInfo } from './dom/getPageInfo'
-import { SimulatorMask } from './mask/SimulatorMask'
 import { patchReact } from './patches/react'

-export { SimulatorMask }
-
 /**
 * Configuration for PageController
 */
@@ -84,7 +81,8 @@ export class PageController extends EventTarget {
 	private lastTimeUpdate = 0

 	/** Visual mask overlay for blocking user interaction during automation */
-	private mask: SimulatorMask | null = null
+	private mask: InstanceType<typeof import('./mask/SimulatorMask').SimulatorMask> | null = null
+	private maskReady: Promise<void> | null = null

 	constructor(config: PageControllerConfig = {}) {
 		super()
@@ -94,10 +92,17 @@ export class PageController extends EventTarget {
 		patchReact(this)

 		if (config.enableMask) {
-			this.mask = new SimulatorMask()
+			this.maskReady = this.initMask()
 		}
 	}

+	/**
+	 * Initialize mask asynchronously (dynamic import to avoid CSS loading in Node)
+	 */
+	private async initMask(): Promise<void> {
+		const { SimulatorMask } = await import('./mask/SimulatorMask')
+		this.mask = new SimulatorMask()
+	}
 	// ======= State Queries =======

 	/**
@@ -366,7 +371,8 @@ export class PageController extends EventTarget {
 	 * Show the visual mask overlay.
 	 * Only works if enableMask was set to true in config.
 	 */
-	showMask(): void {
+	async showMask(): Promise<void> {
+		await this.maskReady
 		this.mask?.show()
 	}

@@ -374,7 +380,8 @@ export class PageController extends EventTarget {
 	 * Hide the visual mask overlay.
 	 * Only works if enableMask was set to true in config.
 	 */
-	hideMask(): void {
+	async hideMask(): Promise<void> {
+		await this.maskReady
 		this.mask?.hide()
 	}

--- a/packages/ui/src/UIState.ts
+++ b/packages/ui/src/UIState.ts
@@ -1,101 +0,0 @@
-/**
- * Agent execution state management
- */
-
-export interface Step {
-	id: string
-	stepNumber: number
-	timestamp: Date
-	type:
-		| 'thinking'
-		| 'tool_executing'
-		| 'completed'
-		| 'error'
-		| 'output'
-		| 'input'
-		| 'retry'
-		| 'observation'
-
-	// Tool execution related
-	toolName?: string
-	toolArgs?: any
-	toolResult?: any
-
-	// Display data
-	displayText: string
-	duration?: number
-}
-
-export type AgentStatus = 'idle' | 'running' | 'completed' | 'error'
-
-export class UIState {
-	private steps: Step[] = []
-	private currentStep: Step | null = null
-	private status: AgentStatus = 'idle'
-	private stepCounter = 0
-
-	addStep(stepData: Omit<Step, 'id' | 'stepNumber' | 'timestamp'>): Step {
-		const step: Step = {
-			id: this.generateId(),
-			stepNumber: ++this.stepCounter,
-			timestamp: new Date(),
-			...stepData,
-		}
-
-		this.steps.push(step)
-		this.currentStep = step
-
-		// Update overall status
-		this.updateStatus(step.type)
-
-		return step
-	}
-
-	updateCurrentStep(updates: Partial<Step>): Step | null {
-		if (!this.currentStep) return null
-
-		Object.assign(this.currentStep, updates)
-		return this.currentStep
-	}
-
-	getCurrentStep(): Step | null {
-		return this.currentStep
-	}
-
-	getAllSteps(): Step[] {
-		return [...this.steps]
-	}
-
-	getStatus(): AgentStatus {
-		return this.status
-	}
-
-	reset(): void {
-		this.steps = []
-		this.currentStep = null
-		this.status = 'idle'
-		this.stepCounter = 0
-	}
-
-	private updateStatus(stepType: Step['type']): void {
-		switch (stepType) {
-			case 'thinking':
-			case 'tool_executing':
-			case 'output':
-			case 'input':
-			case 'retry':
-				this.status = 'running'
-				break
-			case 'completed':
-				this.status = 'completed'
-				break
-			case 'error':
-				this.status = 'error'
-				break
-		}
-	}
-
-	private generateId(): string {
-		return `step_${Date.now()}_${Math.random().toString(36).substring(2, 11)}`
-	}
-}
--- a/packages/ui/src/i18n/locales.ts
+++ b/packages/ui/src/i18n/locales.ts
@@ -22,6 +22,7 @@ const enUS = {
 			selecting: 'Selecting option "{{text}}"...',
 			scrolling: 'Scrolling page...',
 			waiting: 'Waiting {{seconds}} seconds...',
+			askingUser: 'Asking user...',
 			done: 'Task done',
 			clicked: '🖱️ Clicked element [{{index}}]',
 			inputted: '⌨️ Inputted text "{{text}}"',
@@ -68,6 +69,7 @@ const zhCN = {
 			selecting: '正在选择选项 "{{text}}"...',
 			scrolling: '正在滚动页面...',
 			waiting: '等待 {{seconds}} 秒...',
+			askingUser: '正在询问用户...',
 			done: '结束任务',
 			clicked: '🖱️ 已点击元素 [{{index}}]',
 			inputted: '⌨️ 已输入文本 "{{text}}"',
--- a/packages/ui/src/index.ts
+++ b/packages/ui/src/index.ts
@@ -1,3 +1,3 @@
-export { Panel, type PanelConfig, type PanelUpdate } from './Panel'
-export { UIState, type Step, type AgentStatus } from './UIState'
+export { Panel, type PanelConfig } from './panel/Panel'
+export type { AgentActivity, PanelAgentAdapter } from './panel/types'
 export { I18n, type SupportedLanguage, type TranslationKey } from './i18n'
--- a/packages/ui/src/panel/Panel.module.css
+++ b/packages/ui/src/panel/Panel.module.css
@@ -357,6 +357,11 @@
 				background: linear-gradient(135deg, rgba(147, 51, 234, 0.1), rgba(147, 51, 234, 0.05));
 			}

+			&.question {
+				border-left-color: rgb(255, 159, 67);
+				background: linear-gradient(135deg, rgba(255, 159, 67, 0.15), rgba(255, 159, 67, 0.08));
+			}
+
 			/* 突出显示 done 成功结果 */
 			&.doneSuccess {
 				background: linear-gradient(
@@ -439,7 +444,7 @@

 			.historyContent {
 				display: flex;
-				align-items: center;
+				align-items: flex-start;
 				gap: 8px;

 				word-break: break-all;
@@ -453,6 +458,12 @@
 					line-height: 1;
 					transition: all 0.3s ease;
 				}
+
+				.reflectionLines {
+					display: flex;
+					flex-direction: column;
+					gap: 4px;
+				}
 			}

 			.historyMeta {
--- a/packages/ui/src/panel/Panel.ts
+++ b/packages/ui/src/panel/Panel.ts
@@ -1,6 +1,7 @@
-import { type Step, UIState } from './UIState'
-import { I18n, type SupportedLanguage } from './i18n'
-import { escapeHtml, truncate } from './utils'
+import { I18n, type SupportedLanguage } from '../i18n'
+import { truncate } from '../utils'
+import { createCard, createReflectionLines, formatTime } from './cards'
+import type { AgentActivity, PanelAgentAdapter } from './types'

 import styles from './Panel.module.css'

@@ -9,8 +10,6 @@ import styles from './Panel.module.css'
 */
 export interface PanelConfig {
 	language?: SupportedLanguage
-	onExecuteTask: (task: string) => void
-	onStop: () => void
 	/**
 	 * Whether to prompt for next task after task completion
 	 * @default true
@@ -18,24 +17,15 @@ export interface PanelConfig {
 	promptForNextTask?: boolean
 }

-/**
- * Semantic update types - Panel handles i18n internally
- */
-export type PanelUpdate =
-	| { type: 'thinking'; text?: string } // text is optional, defaults to i18n thinking text
-	| { type: 'input'; task: string }
-	| { type: 'question'; question: string }
-	| { type: 'userAnswer'; input: string }
-	| { type: 'retry'; current: number; max: number }
-	| { type: 'error'; message: string }
-	| { type: 'output'; text: string }
-	| { type: 'completed' }
-	| { type: 'toolExecuting'; toolName: string; args: any }
-	| { type: 'toolCompleted'; toolName: string; args: any; result?: string; duration?: number }
-	| { type: 'observation'; content: string }
-
 /**
 * Agent control panel
+ *
+ * Architecture:
+ * - History list: renders directly from agent.history (historical events)
+ * - Header bar: shows activity events (transient state) and agent status
+ *
+ * This separation ensures data consistency - history is the single source of truth
+ * for what has been done, while activity shows what is happening now.
 */
 export class Panel {
 	#wrapper: HTMLElement
@@ -47,9 +37,9 @@ export class Panel {
 	#inputSection: HTMLElement
 	#taskInput: HTMLInputElement

-	#state = new UIState()
-	#isExpanded = false
+	#agent: PanelAgentAdapter
 	#config: PanelConfig
+	#isExpanded = false
 	#i18n: I18n
 	#userAnswerResolver: ((input: string) => void) | null = null
 	#isWaitingForUserAnswer: boolean = false
@@ -57,13 +47,30 @@ export class Panel {
 	#pendingHeaderText: string | null = null
 	#isAnimating = false

+	// Event handlers (bound for removal)
+	#onStatusChange = () => this.#handleStatusChange()
+	#onHistoryChange = () => this.#handleHistoryChange()
+	#onActivity = (e: Event) => this.#handleActivity((e as CustomEvent<AgentActivity>).detail)
+	#onAgentDispose = () => this.dispose()
+
 	get wrapper(): HTMLElement {
 		return this.#wrapper
 	}

-	constructor(config: PanelConfig) {
+	/**
+	 * Create a Panel bound to an agent
+	 * @param agent - Agent instance that implements PanelAgentAdapter
+	 * @param config - Optional panel configuration
+	 */
+	constructor(agent: PanelAgentAdapter, config: PanelConfig = {}) {
+		this.#agent = agent
 		this.#config = config
 		this.#i18n = new I18n(config.language ?? 'en-US')
+
+		// Set up askUser callback on agent
+		this.#agent.onAskUser = (question) => this.#askUser(question)
+
+		// Create UI elements
 		this.#wrapper = this.#createWrapper()
 		this.#indicator = this.#wrapper.querySelector(`.${styles.indicator}`)!
 		this.#statusText = this.#wrapper.querySelector(`.${styles.statusText}`)!
@@ -73,6 +80,12 @@ export class Panel {
 		this.#inputSection = this.#wrapper.querySelector(`.${styles.inputSectionWrapper}`)!
 		this.#taskInput = this.#wrapper.querySelector(`.${styles.taskInput}`)!

+		// Listen to agent events
+		this.#agent.addEventListener('statuschange', this.#onStatusChange)
+		this.#agent.addEventListener('historychange', this.#onHistoryChange)
+		this.#agent.addEventListener('activity', this.#onActivity)
+		this.#agent.addEventListener('dispose', this.#onAgentDispose)
+
 		this.#setupEventListeners()
 		this.#startHeaderUpdateLoop()

@@ -81,24 +94,98 @@ export class Panel {
 		this.hide() // Start hidden
 	}

+	// ========== Agent event handlers ==========
+
+	/** Handle agent status change */
+	#handleStatusChange(): void {
+		const status = this.#agent.status
+
+		// Map agent status to UI indicator type
+		const indicatorType =
+			status === 'running' ? 'thinking' : status === 'idle' ? 'thinking' : status
+		this.#updateStatusIndicator(indicatorType)
+
+		// Show/hide based on status
+		if (status === 'running') {
+			this.show()
+			this.#hideInputArea() // Hide input while running
+		}
+
+		// Handle completion
+		if (status === 'completed' || status === 'error') {
+			if (!this.#isExpanded) {
+				this.#expand()
+			}
+			if (this.#shouldShowInputArea()) {
+				this.#showInputArea()
+			}
+		}
+	}
+
+	/** Handle agent history change - re-render history list from agent.history */
+	#handleHistoryChange(): void {
+		this.#renderHistory()
+	}
+
 	/**
-	 * Ask for user input
+	 * Handle agent activity - transient state for immediate UI feedback
+	 * Activity events are NOT persisted in history, only used for header bar updates
 	 */
-	async askUser(question: string): Promise<string> {
+	#handleActivity(activity: AgentActivity): void {
+		switch (activity.type) {
+			case 'thinking':
+				this.#pendingHeaderText = this.#i18n.t('ui.panel.thinking')
+				this.#updateStatusIndicator('thinking')
+				break
+
+			case 'executing':
+				this.#pendingHeaderText = this.#getToolExecutingText(activity.tool, activity.input)
+				this.#updateStatusIndicator('executing')
+				break
+
+			case 'executed':
+				this.#pendingHeaderText = truncate(activity.output, 50)
+				break
+
+			case 'retrying':
+				this.#pendingHeaderText = `Retrying (${activity.attempt}/${activity.maxAttempts})`
+				this.#updateStatusIndicator('retrying')
+				break
+
+			case 'error':
+				this.#pendingHeaderText = truncate(activity.message, 50)
+				this.#updateStatusIndicator('error')
+				break
+		}
+	}
+
+	/**
+	 * Ask for user input (internal, called by agent via onAskUser)
+	 */
+	#askUser(question: string): Promise<string> {
 		return new Promise((resolve) => {
 			// Set `waiting for user answer` state
 			this.#isWaitingForUserAnswer = true
 			this.#userAnswerResolver = resolve

-			// Update state to `running`
-			this.#updateInternal({
-				type: 'output',
-				displayText: this.#i18n.t('ui.panel.question', { question }),
-			}) // Expand history panel
+			// Expand history panel
 			if (!this.#isExpanded) {
 				this.#expand()
 			}

+			// Add temporary question card so user can see the full question
+			const tempCard = document.createElement('div')
+			tempCard.innerHTML = createCard({
+				icon: '❓',
+				content: `Question: ${question}`,
+				meta: formatTime(this.#config.language ?? 'en-US'),
+				type: 'question',
+			})
+			const cardElement = tempCard.firstElementChild as HTMLElement
+			cardElement.setAttribute('data-temp-card', 'true')
+			this.#historySection.appendChild(cardElement)
+			this.#scrollToBottom()
+
 			this.#showInputArea(this.#i18n.t('ui.panel.userAnswerPrompt'))
 		})
 	}
@@ -119,10 +206,9 @@ export class Panel {
 	}

 	reset(): void {
-		this.#state.reset()
 		this.#statusText.textContent = this.#i18n.t('ui.panel.ready')
 		this.#updateStatusIndicator('thinking')
-		this.#updateHistory()
+		this.#renderHistory()
 		this.#collapse()
 		// Reset user input state
 		this.#isWaitingForUserAnswer = false
@@ -140,17 +226,16 @@ export class Panel {
 	}

 	/**
-	 * Update panel with semantic data - i18n handled internally
-	 */
-	update(data: PanelUpdate): void {
-		const stepData = this.#toStepData(data)
-		this.#updateInternal(stepData)
-	}
-
-	/**
-	 * Dispose panel
+	 * Dispose panel and clean up event listeners
 	 */
 	dispose(): void {
+		// Remove agent event listeners
+		this.#agent.removeEventListener('statuschange', this.#onStatusChange)
+		this.#agent.removeEventListener('historychange', this.#onHistoryChange)
+		this.#agent.removeEventListener('activity', this.#onActivity)
+		this.#agent.removeEventListener('dispose', this.#onAgentDispose)
+
+		// Clean up UI
 		this.#isWaitingForUserAnswer = false
 		this.#stopHeaderUpdateLoop()
 		this.wrapper.remove()
@@ -158,69 +243,21 @@ export class Panel {

 	// ========== Private methods ==========

-	/**
-	 * Convert semantic update to step data with i18n
-	 */
-	#toStepData(data: PanelUpdate): Omit<Step, 'id' | 'stepNumber' | 'timestamp'> {
-		switch (data.type) {
-			case 'thinking':
-				return { type: 'thinking', displayText: data.text ?? this.#i18n.t('ui.panel.thinking') }
-			case 'input':
-				return { type: 'input', displayText: data.task }
-			case 'question':
-				return {
-					type: 'output',
-					displayText: this.#i18n.t('ui.panel.question', { question: data.question }),
-				}
-			case 'userAnswer':
-				return {
-					type: 'input',
-					displayText: this.#i18n.t('ui.panel.userAnswer', { input: data.input }),
-				}
-			case 'retry':
-				return { type: 'retry', displayText: `retry-ing (${data.current} / ${data.max})` }
-			case 'error':
-				return { type: 'error', displayText: data.message }
-			case 'output':
-				return { type: 'output', displayText: data.text }
-			case 'completed':
-				return { type: 'completed', displayText: this.#i18n.t('ui.panel.taskCompleted') }
-			case 'toolExecuting':
-				return {
-					type: 'tool_executing',
-					toolName: data.toolName,
-					toolArgs: data.args,
-					displayText: this.#getToolExecutingText(data.toolName, data.args),
-				}
-			case 'toolCompleted': {
-				const displayText = this.#getToolCompletedText(data.toolName, data.args)
-				if (!displayText) return { type: 'tool_executing', displayText: '' } // will be filtered
-				return {
-					type: 'tool_executing',
-					toolName: data.toolName,
-					toolArgs: data.args,
-					toolResult: data.result,
-					displayText,
-					duration: data.duration,
-				}
-			}
-			case 'observation':
-				return { type: 'observation', displayText: data.content }
-		}
-	}
-
-	#getToolExecutingText(toolName: string, args: any): string {
+	#getToolExecutingText(toolName: string, args: unknown): string {
+		const a = args as Record<string, string | number>
 		switch (toolName) {
 			case 'click_element_by_index':
-				return this.#i18n.t('ui.tools.clicking', { index: args.index })
+				return this.#i18n.t('ui.tools.clicking', { index: a.index })
 			case 'input_text':
-				return this.#i18n.t('ui.tools.inputting', { index: args.index })
+				return this.#i18n.t('ui.tools.inputting', { index: a.index })
 			case 'select_dropdown_option':
-				return this.#i18n.t('ui.tools.selecting', { text: args.text })
+				return this.#i18n.t('ui.tools.selecting', { text: a.text })
 			case 'scroll':
 				return this.#i18n.t('ui.tools.scrolling')
 			case 'wait':
-				return this.#i18n.t('ui.tools.waiting', { seconds: args.seconds })
+				return this.#i18n.t('ui.tools.waiting', { seconds: a.seconds })
+			case 'ask_user':
+				return this.#i18n.t('ui.tools.askingUser')
 			case 'done':
 				return this.#i18n.t('ui.tools.done')
 			default:
@@ -228,67 +265,11 @@ export class Panel {
 		}
 	}

-	#getToolCompletedText(toolName: string, args: any): string | null {
-		switch (toolName) {
-			case 'click_element_by_index':
-				return this.#i18n.t('ui.tools.clicked', { index: args.index })
-			case 'input_text':
-				return this.#i18n.t('ui.tools.inputted', { text: args.text })
-			case 'select_dropdown_option':
-				return this.#i18n.t('ui.tools.selected', { text: args.text })
-			case 'scroll':
-				return this.#i18n.t('ui.tools.scrolled')
-			case 'wait':
-				return this.#i18n.t('ui.tools.waited')
-			case 'done':
-				return null
-			default:
-				return null
-		}
-	}
-
-	/**
-	 * Update status (internal)
-	 */
-	#updateInternal(stepData: Omit<Step, 'id' | 'stepNumber' | 'timestamp'>): void {
-		// Skip empty displayText (filtered toolCompleted for 'done')
-		if (!stepData.displayText) return
-
-		const step = this.#state.addStep(stepData)
-
-		// Queue header text update (will be processed by periodic check)
-		const headerText = truncate(step.displayText, 20)
-		this.#pendingHeaderText = headerText
-
-		this.#updateStatusIndicator(step.type)
-		this.#updateHistory()
-
-		// Auto-expand history after task completion
-		if (step.type === 'completed' || step.type === 'error') {
-			if (!this.#isExpanded) {
-				this.#expand()
-			}
-		}
-
-		// Control input area display based on status
-		if (this.#shouldShowInputArea()) {
-			this.#showInputArea()
-		} else {
-			this.#hideInputArea()
-		}
-	}
-
 	/**
 	 * Stop Agent
 	 */
 	#stopAgent(): void {
-		// Update status display
-		this.#updateInternal({
-			type: 'error',
-			displayText: this.#i18n.t('ui.panel.taskTerminated'),
-		})
-
-		this.#config.onStop()
+		this.#agent.dispose()
 	}

 	/**
@@ -305,7 +286,8 @@ export class Panel {
 			// Handle user input mode
 			this.#handleUserAnswer(input)
 		} else {
-			this.#config.onExecuteTask(input)
+			// Execute task via agent
+			this.#agent.execute(input)
 		}
 	}

@@ -313,10 +295,11 @@ export class Panel {
 	 * Handle user answer
 	 */
 	#handleUserAnswer(input: string): void {
-		// Add user input to history
-		this.#updateInternal({
-			type: 'input',
-			displayText: this.#i18n.t('ui.panel.userAnswer', { input }),
+		// Remove temporary question cards (only direct children for safety)
+		Array.from(this.#historySection.children).forEach((child) => {
+			if (child.getAttribute('data-temp-card') === 'true') {
+				child.remove()
+			}
 		})

 		// Reset state
@@ -357,13 +340,13 @@ export class Panel {
 		// Always show input area if waiting for user input
 		if (this.#isWaitingForUserAnswer) return true

-		const steps = this.#state.getAllSteps()
-		if (steps.length === 0) {
+		const history = this.#agent.history
+		if (history.length === 0) {
 			return true // Initial state
 		}

-		const lastStep = steps[steps.length - 1]
-		const isTaskEnded = lastStep.type === 'completed' || lastStep.type === 'error'
+		const status = this.#agent.status
+		const isTaskEnded = status === 'completed' || status === 'error'

 		// Only show input area after task completion if configured to do so
 		if (isTaskEnded) {
@@ -383,13 +366,12 @@ export class Panel {
 			<div class="${styles.background}"></div>
 			<div class="${styles.historySectionWrapper}">
 				<div class="${styles.historySection}">
-					${this.#createHistoryItem({
-						id: 'placeholder',
-						stepNumber: 0,
-						timestamp: new Date(),
-						type: 'thinking',
-						displayText: this.#i18n.t('ui.panel.waitingPlaceholder'),
-					})}
+					<div class="${styles.historyItem}">
+						<div class="${styles.historyContent}">
+							<span class="${styles.statusIcon}">🧠</span>
+							<span>${this.#i18n.t('ui.panel.waitingPlaceholder')}</span>
+						</div>
+					</div>
 				</div>
 			</div>
 			<div class="${styles.header}">
@@ -544,7 +526,9 @@ export class Panel {
 		}, 150) // Half the duration of fade out animation
 	}

-	#updateStatusIndicator(type: Step['type']): void {
+	#updateStatusIndicator(
+		type: 'thinking' | 'executing' | 'executed' | 'retrying' | 'completed' | 'error'
+	): void {
 		// Clear all status classes
 		this.#indicator.className = styles.indicator

@@ -552,12 +536,6 @@ export class Panel {
 		this.#indicator.classList.add(styles[type])
 	}

-	#updateHistory(): void {
-		const steps = this.#state.getAllSteps()
-		this.#historySection.innerHTML = steps.map((step) => this.#createHistoryItem(step)).join('')
-		this.#scrollToBottom()
-	}
-
 	#scrollToBottom(): void {
 		// Execute in next event loop to ensure DOM update completion
 		setTimeout(() => {
@@ -565,71 +543,107 @@ export class Panel {
 		}, 0)
 	}

-	#createHistoryItem(step: Step): string {
-		const time = step.timestamp.toLocaleTimeString('zh-CN', {
-			hour12: false,
-			hour: '2-digit',
-			minute: '2-digit',
-			second: '2-digit',
-		})
+	/**
+	 * Render history directly from agent.history
+	 *
+	 * Renders:
+	 * 1. Task (first item, from agent.task)
+	 * 2. Reflection cards (evaluation, memory, next_goal)
+	 * 3. Tool execution with output
+	 * 4. Observations
+	 */
+	#renderHistory(): void {
+		const items: string[] = []

-		let typeClass = ''
-		let statusIcon = ''
-
-		// Set styles and icons based on step type
-		if (step.type === 'completed') {
-			// Check if this is a result from done tool
-			if (step.toolName === 'done') {
-				// Judge success or failure based on result
-				const failureKeyword = this.#i18n.t('ui.tools.resultFailure')
-				const errorKeyword = this.#i18n.t('ui.tools.resultError')
-				const isSuccess =
-					!step.toolResult ||
-					(!step.toolResult.includes(failureKeyword) && !step.toolResult.includes(errorKeyword))
-				typeClass = isSuccess ? styles.doneSuccess : styles.doneError
-				statusIcon = isSuccess ? '🎉' : '❌'
-			} else {
-				typeClass = styles.completed
-				statusIcon = '✅'
-			}
-		} else if (step.type === 'error') {
-			typeClass = styles.error
-			statusIcon = '❌'
-		} else if (step.type === 'tool_executing') {
-			statusIcon = '🔨'
-		} else if (step.type === 'output') {
-			typeClass = styles.output
-			statusIcon = '🤖'
-		} else if (step.type === 'input') {
-			typeClass = styles.input
-			statusIcon = '🎯'
-		} else if (step.type === 'retry') {
-			typeClass = styles.retry
-			statusIcon = '🔄'
-		} else if (step.type === 'observation') {
-			typeClass = styles.observation
-			statusIcon = '👁️'
-		} else {
-			statusIcon = '🧠'
+		// 1. Task card (always first)
+		const task = this.#agent.task
+		if (task) {
+			items.push(this.#createTaskCard(task))
 		}

-		const durationText = step.duration ? ` · ${step.duration}ms` : ''
-		const stepLabel = this.#i18n.t('ui.panel.step', {
-			number: step.stepNumber.toString(),
+		// 2. Render each history event
+		const history = this.#agent.history
+		for (let i = 0; i < history.length; i++) {
+			const event = history[i]
+			items.push(...this.#createHistoryCards(event, i + 1))
+		}
+
+		this.#historySection.innerHTML = items.join('')
+		this.#scrollToBottom()
+	}
+
+	#createTaskCard(task: string): string {
+		return createCard({ icon: '🎯', content: task, type: 'input' })
+	}
+
+	/** Create cards for a history event */
+	#createHistoryCards(event: PanelAgentAdapter['history'][number], stepNumber: number): string[] {
+		const cards: string[] = []
+		const time = formatTime(this.#config.language ?? 'en-US')
+		const meta = this.#i18n.t('ui.panel.step', {
+			number: stepNumber.toString(),
 			time,
-			duration: durationText || '', // Explicitly pass empty string to replace template
+			duration: '',
 		})

-		return `
-			<div class="${styles.historyItem} ${typeClass}">
-				<div class="${styles.historyContent}">
-					<span class="${styles.statusIcon}">${statusIcon}</span>
-					<span>${escapeHtml(step.displayText)}</span>
-				</div>
-				<div class="${styles.historyMeta}">
-					${stepLabel}
-				</div>
-			</div>
-		`
+		if (event.type === 'step') {
+			// Reflection card
+			if (event.reflection) {
+				const lines = createReflectionLines(event.reflection)
+				if (lines.length > 0) {
+					cards.push(createCard({ icon: '🧠', content: lines, meta }))
+				}
+			}
+
+			// Action card
+			const action = event.action
+			if (action) {
+				cards.push(...this.#createActionCards(action, meta))
+			}
+		} else if (event.type === 'observation') {
+			cards.push(
+				createCard({ icon: '👁️', content: event.content || '', meta, type: 'observation' })
+			)
+		} else if (event.type === 'user_takeover') {
+			cards.push(createCard({ icon: '👤', content: 'User takeover', meta, type: 'input' }))
+		}
+
+		return cards
+	}
+
+	/** Create cards for an action */
+	#createActionCards(
+		action: { name: string; input: unknown; output: string },
+		meta: string
+	): string[] {
+		const cards: string[] = []
+
+		if (action.name === 'done') {
+			const input = action.input as { text?: string }
+			const text = input.text || action.output || ''
+			if (text) {
+				cards.push(createCard({ icon: '🤖', content: text, meta, type: 'output' }))
+			}
+		} else if (action.name === 'ask_user') {
+			const input = action.input as { question?: string }
+			const answer = action.output.replace(/^User answered:\s*/i, '')
+			cards.push(
+				createCard({
+					icon: '❓',
+					content: `Question: ${input.question || ''}`,
+					meta,
+					type: 'question',
+				})
+			)
+			cards.push(createCard({ icon: '💬', content: `Answer: ${answer}`, meta, type: 'input' }))
+		} else {
+			const toolText = this.#getToolExecutingText(action.name, action.input)
+			cards.push(createCard({ icon: '🔨', content: toolText, meta }))
+			if (action.output?.length > 0) {
+				cards.push(createCard({ icon: '🔨', content: action.output, meta, type: 'output' }))
+			}
+		}
+
+		return cards
 	}
 }
--- a/packages/ui/src/panel/cards.ts
+++ b/packages/ui/src/panel/cards.ts
@@ -0,0 +1,62 @@
+/**
+ * Card HTML generation utilities for Panel
+ */
+import { escapeHtml } from '../utils'
+
+import styles from './Panel.module.css'
+
+type CardType = 'default' | 'input' | 'output' | 'question' | 'observation'
+
+interface CardOptions {
+	icon: string
+	content: string | string[]
+	meta?: string
+	type?: CardType
+}
+
+/** Create a single history card */
+export function createCard({ icon, content, meta, type }: CardOptions): string {
+	const typeClass = type ? styles[type] : ''
+	const contentHtml = Array.isArray(content)
+		? `<div class="${styles.reflectionLines}">${content.join('')}</div>`
+		: `<span>${escapeHtml(content)}</span>`
+
+	return `
+		<div class="${styles.historyItem} ${typeClass}">
+			<div class="${styles.historyContent}">
+				<span class="${styles.statusIcon}">${icon}</span>
+				${contentHtml}
+			</div>
+			${meta ? `<div class="${styles.historyMeta}">${meta}</div>` : ''}
+		</div>
+	`
+}
+
+/** Format timestamp for cards */
+export function formatTime(locale: string = 'en-US'): string {
+	return new Date().toLocaleTimeString(locale, {
+		hour12: false,
+		hour: '2-digit',
+		minute: '2-digit',
+		second: '2-digit',
+	})
+}
+
+/** Create reflection lines from reflection object */
+export function createReflectionLines(reflection: {
+	evaluation_previous_goal?: string
+	memory?: string
+	next_goal?: string
+}): string[] {
+	const lines: string[] = []
+	if (reflection.evaluation_previous_goal) {
+		lines.push(`<div>🔍 ${escapeHtml(reflection.evaluation_previous_goal)}</div>`)
+	}
+	if (reflection.memory) {
+		lines.push(`<div>💾 ${escapeHtml(reflection.memory)}</div>`)
+	}
+	if (reflection.next_goal) {
+		lines.push(`<div>🎯 ${escapeHtml(reflection.next_goal)}</div>`)
+	}
+	return lines
+}
--- a/packages/ui/src/panel/types.ts
+++ b/packages/ui/src/panel/types.ts
@@ -0,0 +1,67 @@
+/**
+ * Agent activity - transient state for immediate UI feedback.
+ *
+ * Unlike historical events (which are persisted), activities are ephemeral
+ * and represent "what the agent is doing right now". UI components should
+ * listen to 'activity' events to show real-time feedback.
+ *
+ * Note: There is no 'idle' activity - absence of activity events means idle.
+ *
+ * Events dispatched: CustomEvent<AgentActivity>
+ */
+export type AgentActivity =
+	| { type: 'thinking' }
+	| { type: 'executing'; tool: string; input: unknown }
+	| { type: 'executed'; tool: string; input: unknown; output: string; duration: number }
+	| { type: 'retrying'; attempt: number; maxAttempts: number }
+	| { type: 'error'; message: string }
+
+/**
+ * Minimal interface that Panel expects from an agent.
+ * Panel does not depend on PageAgent directly - it only requires this interface.
+ * This enables decoupling and allows any agent implementation to work with Panel.
+ *
+ * Events:
+ * - 'statuschange': Agent status changed (idle/running/completed/error)
+ * - 'historychange': Historical events updated (persisted)
+ * - 'activity': Transient activity for immediate UI feedback (thinking/executing/etc)
+ * - 'dispose': Agent is being disposed
+ */
+export interface PanelAgentAdapter extends EventTarget {
+	/** Current agent status */
+	readonly status: 'idle' | 'running' | 'completed' | 'error'
+
+	/** History of agent events */
+	readonly history: readonly {
+		type: 'step' | 'observation' | 'user_takeover' | 'error'
+		/** For 'step' type */
+		reflection?: {
+			evaluation_previous_goal?: string
+			memory?: string
+			next_goal?: string
+		}
+		/** For 'step' type */
+		action?: {
+			name: string
+			input: unknown
+			output: string
+		}
+		/** For 'observation' type */
+		content?: string
+	}[]
+
+	/** Current task being executed */
+	readonly task: string
+
+	/**
+	 * Callback for when agent needs user input.
+	 * Panel will set this to handle user questions via its UI.
+	 */
+	onAskUser?: (question: string) => Promise<string>
+
+	/** Execute a task */
+	execute(task: string): Promise<unknown>
+
+	/** Dispose the agent */
+	dispose(): void
+}
--- a/packages/website/src/components/ui/api-reference.tsx
+++ b/packages/website/src/components/ui/api-reference.tsx
@@ -0,0 +1,167 @@
+/**
+ * API Reference component for displaying TypeScript interface definitions
+ *
+ * Provides a beautiful, readable table for documenting API interfaces
+ */
+import * as React from 'react'
+
+import { cn } from '@/lib/utils'
+
+import { Badge } from './badge'
+
+// ============================================================================
+// Types
+// ============================================================================
+
+export interface PropDefinition {
+	/** Property name */
+	name: string
+	/** TypeScript type (can include generics, unions, etc.) */
+	type: string
+	/** Whether the property is required */
+	required?: boolean
+	/** Default value if any */
+	defaultValue?: string
+	/** Description of the property */
+	description: React.ReactNode
+	/** Mark as experimental/deprecated */
+	status?: 'experimental' | 'deprecated'
+}
+
+export interface APIReferenceProps {
+	/** Title for the API section */
+	title?: string
+	/** Optional description */
+	description?: React.ReactNode
+	/** Property definitions */
+	properties: PropDefinition[]
+	/** Additional CSS classes */
+	className?: string
+}
+
+// ============================================================================
+// Component
+// ============================================================================
+
+export function APIReference({ title, description, properties, className }: APIReferenceProps) {
+	return (
+		<div className={cn('my-6', className)}>
+			{title && (
+				<h3 className="text-lg font-semibold text-gray-900 dark:text-gray-100 mb-2">{title}</h3>
+			)}
+			{description && (
+				<p className="text-sm text-gray-600 dark:text-gray-400 mb-4">{description}</p>
+			)}
+
+			<div className="overflow-hidden rounded-lg border border-gray-200 dark:border-gray-700">
+				<table className="w-full text-sm">
+					<thead>
+						<tr className="bg-gray-50 dark:bg-gray-800/50">
+							<th className="px-4 py-3 text-left font-medium text-gray-600 dark:text-gray-300">
+								Property
+							</th>
+							<th className="px-4 py-3 text-left font-medium text-gray-600 dark:text-gray-300">
+								Type
+							</th>
+							<th className="px-4 py-3 text-left font-medium text-gray-600 dark:text-gray-300 hidden md:table-cell">
+								Default
+							</th>
+							<th className="px-4 py-3 text-left font-medium text-gray-600 dark:text-gray-300">
+								Description
+							</th>
+						</tr>
+					</thead>
+					<tbody className="divide-y divide-gray-100 dark:divide-gray-800">
+						{properties.map((prop) => (
+							<PropRow key={prop.name} {...prop} />
+						))}
+					</tbody>
+				</table>
+			</div>
+		</div>
+	)
+}
+
+function PropRow({ name, type, required, defaultValue, description, status }: PropDefinition) {
+	return (
+		<tr className="bg-white dark:bg-gray-900 hover:bg-gray-50 dark:hover:bg-gray-800/50 transition-colors">
+			{/* Property name */}
+			<td className="px-4 py-3 align-top">
+				<div className="flex items-center gap-2 flex-wrap">
+					<code className="font-mono text-sm font-medium text-indigo-600 dark:text-indigo-400">
+						{name}
+					</code>
+					{required && (
+						<Badge
+							variant="outline"
+							className="text-[10px] px-1.5 py-0 border-red-300 text-red-600 dark:border-red-800 dark:text-red-400"
+						>
+							required
+						</Badge>
+					)}
+					{status === 'experimental' && (
+						<Badge
+							variant="outline"
+							className="text-[10px] px-1.5 py-0 border-amber-300 text-amber-600 dark:border-amber-800 dark:text-amber-400"
+						>
+							experimental
+						</Badge>
+					)}
+					{status === 'deprecated' && (
+						<Badge
+							variant="outline"
+							className="text-[10px] px-1.5 py-0 border-gray-300 text-gray-500 dark:border-gray-700 dark:text-gray-500 line-through"
+						>
+							deprecated
+						</Badge>
+					)}
+				</div>
+			</td>
+
+			{/* Type */}
+			<td className="px-4 py-3 align-top">
+				<code className="font-mono text-xs text-gray-700 dark:text-gray-300 bg-gray-100 dark:bg-gray-800 px-1.5 py-0.5 rounded whitespace-nowrap">
+					{type}
+				</code>
+			</td>
+
+			{/* Default value */}
+			<td className="px-4 py-3 align-top hidden md:table-cell">
+				{defaultValue ? (
+					<code className="font-mono text-xs text-gray-600 dark:text-gray-400">{defaultValue}</code>
+				) : (
+					<span className="text-gray-400 dark:text-gray-600">-</span>
+				)}
+			</td>
+
+			{/* Description */}
+			<td className="px-4 py-3 align-top text-gray-600 dark:text-gray-400">{description}</td>
+		</tr>
+	)
+}
+
+// ============================================================================
+// Utility Components
+// ============================================================================
+
+/** Code inline span for type references in descriptions */
+export function TypeRef({ children }: { children: React.ReactNode }) {
+	return (
+		<code className="font-mono text-xs text-indigo-600 dark:text-indigo-400 bg-indigo-50 dark:bg-indigo-950/50 px-1 py-0.5 rounded">
+			{children}
+		</code>
+	)
+}
+
+/** Section divider for grouping related APIs */
+export function APIDivider({ title }: { title: string }) {
+	return (
+		<div className="flex items-center gap-4 my-8">
+			<div className="h-px flex-1 bg-gradient-to-r from-transparent via-gray-200 dark:via-gray-700 to-transparent" />
+			<span className="text-xs font-medium uppercase tracking-wider text-gray-500 dark:text-gray-400">
+				{title}
+			</span>
+			<div className="h-px flex-1 bg-gradient-to-r from-transparent via-gray-200 dark:via-gray-700 to-transparent" />
+		</div>
+	)
+}
--- a/packages/website/src/i18n/locales/en-US/common.ts
+++ b/packages/website/src/i18n/locales/en-US/common.ts
@@ -24,6 +24,7 @@ export default {
 		introduction: 'Introduction',
 		features: 'Features',
 		integration: 'Integration',
+		advanced: 'Advanced',
 		overview: 'Overview',
 		quick_start: 'Quick Start',
 		limitations: 'Limitations',
@@ -32,9 +33,10 @@ export default {
 		knowledge_injection: 'Instructions',
 		data_masking: 'Data Masking',
 		cdn_setup: 'CDN Setup',
-		configuration: 'Configuration',
 		best_practices: 'Best Practices',
 		third_party_agent: 'Third-party Agent',
 		security_permissions: 'Security & Permissions',
+		page_agent: 'PageAgent',
+		page_agent_core: 'PageAgentCore',
 	},
 }
--- a/packages/website/src/i18n/locales/zh-CN/common.ts
+++ b/packages/website/src/i18n/locales/zh-CN/common.ts
@@ -23,6 +23,7 @@ export default {
 		introduction: '介绍',
 		features: '功能特性',
 		integration: '集成指南',
+		advanced: '高级',
 		overview: '概览',
 		quick_start: '快速开始',
 		limitations: '使用限制',
@@ -31,9 +32,10 @@ export default {
 		knowledge_injection: '知识注入',
 		data_masking: '数据脱敏',
 		cdn_setup: 'CDN 引入',
-		configuration: '配置选项',
 		best_practices: '最佳实践',
 		third_party_agent: '接入第三方 Agent',
 		security_permissions: '安全与权限',
+		page_agent: 'PageAgent',
+		page_agent_core: 'PageAgentCore',
 	},
 }
--- a/packages/website/src/pages/docs/Layout.tsx
+++ b/packages/website/src/pages/docs/Layout.tsx
@@ -41,7 +41,6 @@ export default function DocsLayout({ children }: DocsLayoutProps) {
 		{
 			title: t('nav.integration'),
 			items: [
-				{ title: t('nav.configuration'), path: '/integration/configuration' },
 				{ title: t('nav.third_party_agent'), path: '/integration/third-party-agent' },
 				{ title: t('nav.cdn_setup'), path: '/integration/cdn-setup' },
 				{
@@ -51,6 +50,13 @@ export default function DocsLayout({ children }: DocsLayoutProps) {
 				{ title: '🚧 ' + t('nav.best_practices'), path: '/integration/best-practices' },
 			],
 		},
+		{
+			title: t('nav.advanced'),
+			items: [
+				{ title: t('nav.page_agent'), path: '/advanced/page-agent' },
+				{ title: t('nav.page_agent_core'), path: '/advanced/page-agent-core' },
+			],
+		},
 	]

 	return (
--- a/packages/website/src/pages/docs/advanced/page-agent-core/page.tsx
+++ b/packages/website/src/pages/docs/advanced/page-agent-core/page.tsx
@@ -0,0 +1,514 @@
+import { useTranslation } from 'react-i18next'
+
+import CodeEditor from '@/components/CodeEditor'
+import { APIDivider, APIReference, TypeRef } from '@/components/ui/api-reference'
+
+export default function PageAgentCoreDocs() {
+	const { i18n } = useTranslation()
+	const isZh = i18n.language === 'zh-CN'
+
+	return (
+		<div>
+			<h1 className="text-4xl font-bold mb-6">PageAgentCore</h1>
+
+			<p className="text-xl text-gray-600 dark:text-gray-300 mb-8 leading-relaxed">
+				{isZh
+					? 'PageAgentCore 是不带 UI 的核心 Agent 类。用于需要自定义 UI 或无头运行的场景。'
+					: 'PageAgentCore is the core Agent class without UI. Use it for custom UI or headless scenarios.'}
+			</p>
+
+			{/* When to use */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">
+					{isZh ? '何时使用 PageAgentCore' : 'When to Use PageAgentCore'}
+				</h2>
+				<ul className="list-disc list-inside text-gray-600 dark:text-gray-400 space-y-2">
+					<li>{isZh ? '需要自定义 UI 界面' : 'Need a custom UI interface'}</li>
+					<li>{isZh ? '在自动化测试中无头运行' : 'Running headless in automated tests'}</li>
+					<li>
+						{isZh
+							? '在非浏览器环境运行（需自定义 PageController）'
+							: 'Running in non-browser environments (requires custom PageController)'}
+					</li>
+					<li>
+						{isZh
+							? '将 PageAgent 嵌入其他 Agent 系统'
+							: 'Embedding PageAgent in other agent systems'}
+					</li>
+				</ul>
+			</section>
+
+			{/* Basic Usage */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '基本用法' : 'Basic Usage'}</h2>
+				<CodeEditor
+					language="typescript"
+					code={`import { PageAgentCore } from '@page-agent/core'
+import { PageController } from '@page-agent/page-controller'
+
+const agent = new PageAgentCore({
+  pageController: new PageController({ enableMask: true }),
+  baseURL: 'https://api.openai.com/v1',
+  apiKey: 'your-api-key',
+  model: 'gpt-5.2',
+  language: 'en-US',
+})
+
+// Listen to events for UI display
+
+agent.addEventListener('statuschange', () => {
+  console.log('Status:', agent.status)
+})
+
+agent.addEventListener('historychange', () => {
+  console.log('History:', agent.history)
+})
+
+agent.addEventListener('activity', (e) => {
+  const activity = (e as CustomEvent).detail
+  console.log('Activity:', activity.type)
+})
+
+// Execute task
+const result = await agent.execute('Fill in the form with test data')`}
+				/>
+			</section>
+
+			<APIDivider title={isZh ? '配置' : 'Configuration'} />
+
+			{/* LLM Configuration */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">LLMConfig</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh
+						? '配置与大语言模型的连接参数。支持 OpenAI 兼容的 API。'
+						: 'Configure connection parameters for the language model. Supports OpenAI-compatible APIs.'}
+				</p>
+				<APIReference
+					properties={[
+						{
+							name: 'baseURL',
+							type: 'string',
+							required: true,
+							description: isZh
+								? 'LLM API 的基础 URL（如 https://api.openai.com/v1）'
+								: 'Base URL of the LLM API (e.g., https://api.openai.com/v1)',
+						},
+						{
+							name: 'apiKey',
+							type: 'string',
+							required: true,
+							description: isZh ? 'API 密钥' : 'API key for authentication',
+						},
+						{
+							name: 'model',
+							type: 'string',
+							required: true,
+							description: isZh
+								? '模型名称（如 gpt-4o, claude-3.5-sonnet）'
+								: 'Model name (e.g., gpt-4o, claude-3.5-sonnet)',
+						},
+						{
+							name: 'temperature',
+							type: 'number',
+							defaultValue: '0',
+							description: isZh
+								? '模型温度参数，控制输出随机性'
+								: 'Model temperature, controls output randomness',
+						},
+						{
+							name: 'maxRetries',
+							type: 'number',
+							defaultValue: '3',
+							description: isZh ? 'API 调用失败时的最大重试次数' : 'Maximum retries on API failure',
+						},
+						{
+							name: 'customFetch',
+							type: 'typeof fetch',
+							description: isZh
+								? '自定义 fetch 函数，用于定制 headers、credentials、代理等'
+								: 'Custom fetch function for customizing headers, credentials, proxy, etc.',
+						},
+					]}
+				/>
+			</section>
+
+			{/* Agent Configuration */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">AgentConfig</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh
+						? '配置 Agent 的行为、生命周期钩子和扩展能力。'
+						: 'Configure agent behavior, lifecycle hooks, and extension capabilities.'}
+				</p>
+				<APIReference
+					properties={[
+						{
+							name: 'language',
+							type: "'en-US' | 'zh-CN'",
+							defaultValue: "'en-US'",
+							description: isZh ? 'Agent 输出语言' : 'Agent output language',
+						},
+						{
+							name: 'customTools',
+							type: 'Record<string, PageAgentTool | null>',
+							status: 'experimental',
+							description: isZh
+								? '自定义工具，可扩展或覆盖内置工具。设为 null 可移除工具。'
+								: 'Custom tools to extend or override built-in tools. Set to null to remove a tool.',
+						},
+						{
+							name: 'instructions',
+							type: 'InstructionsConfig',
+							description: isZh
+								? '指导 Agent 行为的指令配置'
+								: 'Instructions to guide agent behavior',
+						},
+						{
+							name: 'transformPageContent',
+							type: '(content: string) => string | Promise<string>',
+							description: isZh
+								? '发送给 LLM 前转换页面内容，可用于数据脱敏'
+								: 'Transform page content before sending to LLM, useful for data masking',
+						},
+						{
+							name: 'experimentalScriptExecutionTool',
+							type: 'boolean',
+							defaultValue: 'false',
+							status: 'experimental',
+							description: isZh
+								? '启用实验性 JavaScript 执行工具'
+								: 'Enable experimental JavaScript execution tool',
+						},
+					]}
+				/>
+
+				<h3 className="text-lg font-semibold mt-6 mb-3">InstructionsConfig</h3>
+				<APIReference
+					properties={[
+						{
+							name: 'system',
+							type: 'string',
+							description: isZh
+								? '全局系统级指令，应用于所有任务'
+								: 'Global system-level instructions, applied to all tasks',
+						},
+						{
+							name: 'getPageInstructions',
+							type: '(url: string) => string | undefined | null',
+							description: isZh
+								? '动态页面级指令回调，在每个步骤前调用'
+								: 'Dynamic page-level instructions callback, called before each step',
+						},
+					]}
+				/>
+			</section>
+
+			{/* Lifecycle Hooks */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '生命周期钩子' : 'Lifecycle Hooks'}</h2>
+				<APIReference
+					properties={[
+						{
+							name: 'onBeforeStep',
+							type: '(stepCnt: number) => void | Promise<void>',
+							description: isZh ? '每个步骤执行前调用' : 'Called before each step execution',
+							status: 'experimental',
+						},
+						{
+							name: 'onAfterStep',
+							type: '(history: HistoricalEvent[]) => void | Promise<void>',
+							description: isZh ? '每个步骤执行后调用' : 'Called after each step execution',
+							status: 'experimental',
+						},
+						{
+							name: 'onBeforeTask',
+							type: '() => void | Promise<void>',
+							description: isZh ? '任务开始前调用' : 'Called before task starts',
+							status: 'experimental',
+						},
+						{
+							name: 'onAfterTask',
+							type: '(result: ExecutionResult) => void | Promise<void>',
+							description: isZh ? '任务结束后调用' : 'Called after task ends',
+							status: 'experimental',
+						},
+						{
+							name: 'onDispose',
+							type: '(reason?: string) => void',
+							description: isZh ? 'Agent 销毁时调用' : 'Called when agent is disposed',
+							status: 'experimental',
+						},
+					]}
+				/>
+			</section>
+
+			{/* PageController Configuration */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">PageControllerConfig</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh
+						? '配置 DOM 提取、元素交互和视觉反馈。'
+						: 'Configure DOM extraction, element interaction, and visual feedback.'}
+				</p>
+				<APIReference
+					properties={[
+						{
+							name: 'pageController',
+							type: 'PageController',
+							status: 'experimental',
+							description: isZh
+								? '自定义 PageController 实例。如不提供，将创建默认实例。'
+								: 'Custom PageController instance. If not provided, a default one will be created.',
+						},
+						{
+							name: 'enableMask',
+							type: 'boolean',
+							defaultValue: 'true',
+							description: isZh
+								? '启用视觉遮罩覆盖层，阻止用户在自动化期间操作'
+								: 'Enable visual mask overlay that blocks user interaction during automation',
+						},
+						{
+							name: 'viewportExpansion',
+							type: 'number',
+							defaultValue: '0',
+							description: isZh
+								? '视口扩展像素数，-1 表示提取整个页面'
+								: 'Viewport expansion in pixels, -1 means extract entire page',
+						},
+						{
+							name: 'interactiveBlacklist',
+							type: '(Element | (() => Element))[]',
+							description: isZh ? '要排除的交互元素列表' : 'Elements to exclude from interaction',
+						},
+						{
+							name: 'interactiveWhitelist',
+							type: '(Element | (() => Element))[]',
+							description: isZh
+								? '要强制包含的交互元素列表'
+								: 'Elements to force include for interaction',
+						},
+						{
+							name: 'include_attributes',
+							type: 'string[]',
+							description: isZh
+								? '在 DOM 提取中包含的额外属性'
+								: 'Additional attributes to include in DOM extraction',
+						},
+					]}
+				/>
+			</section>
+
+			<APIDivider title={isZh ? '属性与方法' : 'Properties & Methods'} />
+
+			{/* Properties */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '属性' : 'Properties'}</h2>
+				<APIReference
+					properties={[
+						{
+							name: 'status',
+							type: "'idle' | 'running' | 'completed' | 'error'",
+							description: isZh ? '当前 Agent 执行状态' : 'Current agent execution status',
+						},
+						{
+							name: 'history',
+							type: 'HistoricalEvent[]',
+							description: isZh
+								? '历史事件数组，构成 Agent 的记忆'
+								: 'Array of historical events, forms agent memory',
+						},
+						{
+							name: 'task',
+							type: 'string',
+							description: isZh ? '当前正在执行的任务' : 'Current task being executed',
+						},
+						{
+							name: 'pageController',
+							type: 'PageController',
+							description: isZh
+								? 'PageController 实例，用于 DOM 操作'
+								: 'PageController instance for DOM operations',
+						},
+						{
+							name: 'tools',
+							type: 'Map<string, PageAgentTool>',
+							description: isZh ? '可用工具的 Map' : 'Map of available tools',
+						},
+						{
+							name: 'onAskUser',
+							type: '(question: string) => Promise<string>',
+							description: isZh
+								? 'Agent 需要用户输入时的回调。未设置则禁用 ask_user 工具。'
+								: 'Callback when agent needs user input. If not set, ask_user tool is disabled.',
+						},
+					]}
+				/>
+			</section>
+
+			{/* Methods */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '方法' : 'Methods'}</h2>
+				<APIReference
+					properties={[
+						{
+							name: 'execute(task: string)',
+							type: 'Promise<ExecutionResult>',
+							description: isZh
+								? '执行任务并返回结果。包含 success、data 和 history 字段。'
+								: 'Execute a task and return result. Contains success, data, and history fields.',
+						},
+						{
+							name: 'pushObservation(content: string)',
+							type: 'void',
+							description: isZh
+								? '向历史流推送一个观察事件，会在下一步时被 LLM 看到'
+								: 'Push an observation to history stream, will be seen by LLM in next step',
+						},
+						{
+							name: 'emitActivity(activity: AgentActivity)',
+							type: 'void',
+							description: isZh
+								? '发出活动事件用于 UI 反馈'
+								: 'Emit activity event for UI feedback',
+						},
+						{
+							name: 'dispose(reason?: string)',
+							type: 'void',
+							description: isZh
+								? '销毁 Agent 并清理资源'
+								: 'Dispose the agent and clean up resources',
+						},
+					]}
+				/>
+			</section>
+
+			{/* Events */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '事件' : 'Events'}</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh ? (
+						<>
+							PageAgentCore 继承自 <TypeRef>EventTarget</TypeRef>，提供以下事件：
+						</>
+					) : (
+						<>
+							PageAgentCore extends <TypeRef>EventTarget</TypeRef> and provides the following
+							events:
+						</>
+					)}
+				</p>
+				<APIReference
+					properties={[
+						{
+							name: 'statuschange',
+							type: 'Event',
+							description: isZh
+								? 'Agent 状态变化时触发 (idle → running → completed/error)'
+								: 'Fired when agent status changes (idle → running → completed/error)',
+						},
+						{
+							name: 'historychange',
+							type: 'Event',
+							description: isZh
+								? '历史事件更新时触发（持久化事件，构成 Agent 记忆）'
+								: 'Fired when history events are updated (persistent, part of agent memory)',
+						},
+						{
+							name: 'activity',
+							type: 'CustomEvent<AgentActivity>',
+							description: isZh
+								? '实时活动反馈（短暂状态，仅用于 UI）。类型包括：thinking, executing, executed, retrying, error'
+								: 'Real-time activity feedback (transient, UI only). Types: thinking, executing, executed, retrying, error',
+						},
+						{
+							name: 'dispose',
+							type: 'Event',
+							description: isZh ? 'Agent 被销毁时触发' : 'Fired when agent is disposed',
+						},
+					]}
+				/>
+			</section>
+
+			<APIDivider title={isZh ? '类型定义' : 'Type Definitions'} />
+
+			{/* ExecutionResult */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">ExecutionResult</h2>
+				<CodeEditor
+					language="typescript"
+					code={`interface ExecutionResult {
+  /** Whether the task completed successfully */
+  success: boolean
+  /** Result description from the agent */
+  data: string
+  /** Full execution history */
+  history: HistoricalEvent[]
+}`}
+				/>
+			</section>
+
+			{/* AgentActivity */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">AgentActivity</h2>
+				<CodeEditor
+					language="typescript"
+					code={`type AgentActivity =
+  | { type: 'thinking' }
+  | { type: 'executing'; tool: string; input: unknown }
+  | { type: 'executed'; tool: string; input: unknown; output: string; duration: number }
+  | { type: 'retrying'; attempt: number; maxAttempts: number }
+  | { type: 'error'; message: string }`}
+				/>
+			</section>
+
+			<APIDivider title={isZh ? '无头模式' : 'Headless Mode'} />
+
+			{/* Headless Usage */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '无头模式' : 'Headless Mode'}</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh
+						? '在非 DOM 环境中，你必须实现自定义的 PageController（例如远程操作页面或 Puppeteer）。'
+						: 'In non-DOM environments, you must implement a custom PageController (e.g., remote page control or Puppeteer).'}
+				</p>
+				<CodeEditor
+					language="typescript"
+					code={`import { PageAgentCore } from '@page-agent/core'
+import type { PageController } from '@page-agent/page-controller'
+
+class MyRemotePageController implements PageController {
+  // Implement required methods for DOM extraction and interaction
+}
+
+const agent = new PageAgentCore({
+  pageController: new MyRemotePageController(),
+  baseURL: 'https://api.openai.com/v1',
+  apiKey: 'your-api-key',
+  model: 'gpt-5.2',
+  language: 'en-US',
+})
+
+// Listen to events for UI display
+
+agent.addEventListener('statuschange', () => {
+  console.log('Status:', agent.status)
+})
+
+agent.addEventListener('historychange', () => {
+  console.log('History:', agent.history)
+})
+
+agent.addEventListener('activity', (e) => {
+  const activity = (e as CustomEvent).detail
+  console.log('Activity:', activity.type)
+})
+
+// Execute task
+const result = await agent.execute('Fill in the form with test data')`}
+				/>
+			</section>
+		</div>
+	)
+}
--- a/packages/website/src/pages/docs/advanced/page-agent/page.tsx
+++ b/packages/website/src/pages/docs/advanced/page-agent/page.tsx
@@ -0,0 +1,246 @@
+import { useTranslation } from 'react-i18next'
+import { Link } from 'wouter'
+
+import CodeEditor from '@/components/CodeEditor'
+import { APIReference, TypeRef } from '@/components/ui/api-reference'
+
+export default function PageAgentDocs() {
+	const { i18n } = useTranslation()
+	const isZh = i18n.language === 'zh-CN'
+
+	return (
+		<div>
+			<h1 className="text-4xl font-bold mb-6">PageAgent</h1>
+
+			<p className="text-xl text-gray-600 dark:text-gray-300 mb-8 leading-relaxed">
+				{isZh
+					? 'PageAgent 是带有内置 UI 面板的完整 Agent 类。它继承自 PageAgentCore，并自动创建交互面板。'
+					: 'PageAgent is the complete Agent class with built-in UI panel. It extends PageAgentCore and automatically creates an interactive panel.'}
+			</p>
+
+			{/* When to use */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">
+					{isZh ? '何时使用 PageAgent' : 'When to Use PageAgent'}
+				</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh
+						? '在大多数场景下，你应该使用 PageAgent。它提供了开箱即用的完整体验：'
+						: 'In most cases, you should use PageAgent. It provides a complete out-of-the-box experience:'}
+				</p>
+				<ul className="list-disc list-inside text-gray-600 dark:text-gray-400 space-y-2 mb-6">
+					<li>
+						{isZh
+							? '内置 UI 面板，显示任务进度、Agent 思考过程和操作结果'
+							: 'Built-in UI panel showing task progress, agent thinking, and action results'}
+					</li>
+					<li>
+						{isZh
+							? '支持 ask_user 工具，Agent 可以向用户提问'
+							: 'Supports ask_user tool for agent to ask questions to users'}
+					</li>
+				</ul>
+			</section>
+
+			{/* Basic Usage */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '基本用法' : 'Basic Usage'}</h2>
+				<CodeEditor
+					language="typescript"
+					code={`import { PageAgent } from 'page-agent'
+
+const agent = new PageAgent({
+  // LLM Configuration (required)
+  baseURL: 'https://api.openai.com/v1',
+  apiKey: 'your-api-key',
+  model: 'gpt-4o',
+  
+  // Optional settings
+  language: 'en-US',
+})
+
+// Execute a task
+const result = await agent.execute('Click the login button')
+
+console.log(result.success) // true or false
+console.log(result.data)    // Task result description
+console.log(result.history) // Full execution history`}
+				/>
+			</section>
+
+			{/* Class Definition */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '类定义' : 'Class Definition'}</h2>
+				<CodeEditor
+					language="typescript"
+					code={`class PageAgent extends PageAgentCore {
+  /** The UI panel instance */
+  panel: Panel
+
+  constructor(config: PageAgentConfig)
+}`}
+				/>
+				<p className="text-gray-600 dark:text-gray-400 mt-4">
+					{isZh ? (
+						<>
+							PageAgent 继承自{' '}
+							<Link
+								href="/advanced/page-agent-core"
+								className="text-blue-600 dark:text-blue-400 hover:underline"
+							>
+								PageAgentCore
+							</Link>
+							，所有核心方法和事件都可用。详细的 API 参考请查看 PageAgentCore 文档。
+						</>
+					) : (
+						<>
+							PageAgent extends{' '}
+							<Link
+								href="/advanced/page-agent-core"
+								className="text-blue-600 dark:text-blue-400 hover:underline"
+							>
+								PageAgentCore
+							</Link>
+							. All core methods and events are available. See PageAgentCore docs for detailed API
+							reference.
+						</>
+					)}
+				</p>
+			</section>
+
+			{/* Configuration */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? '配置' : 'Configuration'}</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh
+						? 'PageAgent 使用与 PageAgentCore 相同的配置接口。'
+						: 'PageAgent uses the same configuration interface as PageAgentCore.'}
+				</p>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh ? (
+						<>
+							完整配置请参考{' '}
+							<Link
+								href="/advanced/page-agent-core"
+								className="text-blue-600 dark:text-blue-400 hover:underline"
+							>
+								PageAgentCore 配置文档
+							</Link>
+							。
+						</>
+					) : (
+						<>
+							See{' '}
+							<Link
+								href="/advanced/page-agent-core"
+								className="text-blue-600 dark:text-blue-400 hover:underline"
+							>
+								PageAgentCore configuration docs
+							</Link>{' '}
+							for complete reference.
+						</>
+					)}
+				</p>
+			</section>
+
+			{/* Panel Property */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">{isZh ? 'Panel 属性' : 'Panel Property'}</h2>
+				<p className="text-gray-600 dark:text-gray-400 mb-4">
+					{isZh
+						? 'PageAgent 自动创建一个 Panel 实例。你可以通过 panel 属性访问它来控制 UI：'
+						: 'PageAgent automatically creates a Panel instance. You can access it via the panel property to control the UI:'}
+				</p>
+
+				<APIReference
+					properties={[
+						{
+							name: 'panel',
+							type: 'Panel',
+							required: true,
+							description: isZh
+								? '内置的 UI 面板实例，用于显示任务进度和接收用户输入。'
+								: 'The built-in UI panel instance for displaying task progress and receiving user input.',
+						},
+					]}
+				/>
+
+				<h3 className="text-lg font-semibold mt-6 mb-3">{isZh ? 'Panel 方法' : 'Panel Methods'}</h3>
+				<CodeEditor
+					language="typescript"
+					code={`// Show/hide the panel
+agent.panel.show()
+agent.panel.hide()
+
+// Expand/collapse history view
+agent.panel.expand()
+agent.panel.collapse()
+
+// Reset panel state
+agent.panel.reset()
+
+// Dispose panel (called automatically when agent disposes)
+agent.panel.dispose()`}
+				/>
+			</section>
+
+			{/* Comparison with PageAgentCore */}
+			<section className="mb-10">
+				<h2 className="text-2xl font-semibold mb-4">
+					{isZh ? 'PageAgent vs PageAgentCore' : 'PageAgent vs PageAgentCore'}
+				</h2>
+				<div className="overflow-hidden rounded-lg border border-gray-200 dark:border-gray-700">
+					<table className="w-full text-sm">
+						<thead>
+							<tr className="bg-gray-50 dark:bg-gray-800/50">
+								<th className="px-4 py-3 text-left font-medium text-gray-600 dark:text-gray-300">
+									{isZh ? '特性' : 'Feature'}
+								</th>
+								<th className="px-4 py-3 text-center font-medium text-gray-600 dark:text-gray-300">
+									PageAgent
+								</th>
+								<th className="px-4 py-3 text-center font-medium text-gray-600 dark:text-gray-300">
+									PageAgentCore
+								</th>
+							</tr>
+						</thead>
+						<tbody className="divide-y divide-gray-100 dark:divide-gray-800">
+							<tr className="bg-white dark:bg-gray-900">
+								<td className="px-4 py-3 text-gray-600 dark:text-gray-400">
+									{isZh ? 'UI 面板' : 'UI Panel'}
+								</td>
+								<td className="px-4 py-3 text-center text-green-600 dark:text-green-400">✓</td>
+								<td className="px-4 py-3 text-center text-gray-400 dark:text-gray-600">-</td>
+							</tr>
+							<tr className="bg-white dark:bg-gray-900">
+								<td className="px-4 py-3 text-gray-600 dark:text-gray-400">
+									{isZh ? 'Headless 模式' : 'Headless Mode'}
+								</td>
+								<td className="px-4 py-3 text-center text-gray-400 dark:text-gray-600">-</td>
+								<td className="px-4 py-3 text-center text-green-600 dark:text-green-400">✓</td>
+							</tr>
+							<tr className="bg-white dark:bg-gray-900">
+								<td className="px-4 py-3 text-gray-600 dark:text-gray-400">
+									{isZh ? '自定义 PageController' : 'Custom PageController'}
+								</td>
+								<td className="px-4 py-3 text-center text-green-600 dark:text-green-400">✓</td>
+								<td className="px-4 py-3 text-center text-green-600 dark:text-green-400">✓</td>
+							</tr>
+							<tr className="bg-white dark:bg-gray-900">
+								<td className="px-4 py-3 text-gray-600 dark:text-gray-400">
+									{isZh ? '适用场景' : 'Use Case'}
+								</td>
+								<td className="px-4 py-3 text-center text-gray-600 dark:text-gray-400">
+									{isZh ? '网页集成' : 'Web integration'}
+								</td>
+								<td className="px-4 py-3 text-center text-gray-600 dark:text-gray-400">
+									{isZh ? '自定义 UI / 无头' : 'Custom UI / Headless'}
+								</td>
+							</tr>
+						</tbody>
+					</table>
+				</div>
+			</section>
+		</div>
+	)
+}
--- a/packages/website/src/pages/docs/index.tsx
+++ b/packages/website/src/pages/docs/index.tsx
@@ -3,6 +3,9 @@ import { Route, Switch } from 'wouter'

 import Header from '../../components/Header'
 import DocsLayout from './Layout'
+import PageAgentCoreDocs from './advanced/page-agent-core/page'
+// Advanced
+import PageAgentDocs from './advanced/page-agent/page'
 import Instructions from './features/custom-instructions/page'
 // Features
 import CustomTools from './features/custom-tools/page'
@@ -11,7 +14,6 @@ import Models from './features/models/page'
 import BestPractices from './integration/best-practices/page'
 // Integration
 import CdnSetup from './integration/cdn-setup/page'
-import Configuration from './integration/configuration/page'
 import SecurityPermissions from './integration/security-permissions/page'
 import ThirdPartyAgent from './integration/third-party-agent/page'
 import Limitations from './introduction/limitations/page'
@@ -83,11 +85,6 @@ export default function DocsRouter() {
 					<SecurityPermissions />
 				</DocsPage>
 			</Route>
-			<Route path="/integration/configuration">
-				<DocsPage>
-					<Configuration />
-				</DocsPage>
-			</Route>
 			<Route path="/integration/best-practices">
 				<DocsPage>
 					<BestPractices />
@@ -99,6 +96,18 @@ export default function DocsRouter() {
 				</DocsPage>
 			</Route>

+			{/* Advanced */}
+			<Route path="/advanced/page-agent">
+				<DocsPage>
+					<PageAgentDocs />
+				</DocsPage>
+			</Route>
+			<Route path="/advanced/page-agent-core">
+				<DocsPage>
+					<PageAgentCoreDocs />
+				</DocsPage>
+			</Route>
+
 			{/* Default redirect or 404 */}
 			<Route path="/docs">
 				<DocsPage>
--- a/packages/website/src/pages/docs/integration/configuration/page.tsx
+++ b/packages/website/src/pages/docs/integration/configuration/page.tsx
@@ -1,199 +0,0 @@
-import { useTranslation } from 'react-i18next'
-
-import CodeEditor from '@/components/CodeEditor'
-
-export default function Configuration() {
-	const { i18n } = useTranslation()
-	const isZh = i18n.language === 'zh-CN'
-
-	return (
-		<div>
-			<h1 className="text-4xl font-bold mb-6">{isZh ? '配置选项' : 'Configuration'}</h1>
-
-			<p className="text-xl text-gray-600 dark:text-gray-300 mb-8 leading-relaxed">
-				{isZh
-					? 'PageAgent 的完整配置接口定义。'
-					: 'Complete configuration interface for PageAgent.'}
-			</p>
-
-			{/* LLM Configuration */}
-			<section className="mb-10">
-				<h2 className="text-2xl font-semibold mb-4">{isZh ? 'LLM 配置' : 'LLM Configuration'}</h2>
-				<p className="text-gray-600 dark:text-gray-400 mb-4">
-					{isZh
-						? '配置与大语言模型的连接参数。'
-						: 'Configure connection parameters for the language model.'}
-				</p>
-				<CodeEditor
-					className="mb-4"
-					language="typescript"
-					code={`interface LLMConfig {
-  baseURL: string
-  apiKey: string
-  model: string
-
-  temperature?: number
-  maxRetries?: number
-
-  /**
-   * Custom fetch function for LLM API requests.
-   * Use this to customize headers, credentials, proxy, etc.
-   * The response should follow OpenAI API format.
-   */
-  customFetch?: typeof globalThis.fetch
-}`}
-				/>
-			</section>
-
-			{/* Agent Configuration */}
-			<section className="mb-10">
-				<h2 className="text-2xl font-semibold mb-4">
-					{isZh ? 'Agent 配置' : 'Agent Configuration'}
-				</h2>
-				<p className="text-gray-600 dark:text-gray-400 mb-4">
-					{isZh
-						? '配置 Agent 的行为、生命周期钩子和扩展能力。'
-						: 'Configure agent behavior, lifecycle hooks, and extension capabilities.'}
-				</p>
-				<CodeEditor
-					className="mb-4"
-					language="typescript"
-					code={`interface AgentConfig {
-  language?: 'en-US' | 'zh-CN'
-
-  /**
-   * Whether to prompt for next task after task completion
-   * @default true
-   */
-  promptForNextTask?: boolean
-
-  /**
-   * Enable the UI panel for visual feedback and user interaction
-   * When disabled, the panel will not be created and all UI operations will be skipped.
-   * Useful for automated testing or when integrating PageAgent as a library.
-   * @default true
-   */
-  enablePanel?: boolean
-
-  /**
-   * Enable the ask_user tool for agent to ask questions
-   * When disabled, the agent cannot ask user questions during execution.
-   * @default true
-   */
-  enableAskUser?: boolean
-
-  /** Custom tools to extend or override built-in tools */
-  customTools?: Record<string, PageAgentTool | null>
-
-  /** Instructions to guide the agent's behavior */
-  instructions?: {
-    /** Global system-level instructions, applied to all tasks */
-    system?: string
-
-    /** Dynamic page-level instructions callback */
-    getPageInstructions?: (url: string) => string | undefined | null
-  }
-
-  // Lifecycle hooks (with \`this\` bound to PageAgent instance)
-  onBeforeStep?: (this: PageAgent, stepCnt: number) => Promise<void> | void
-  onAfterStep?: (this: PageAgent, stepCnt: number, history: HistoryEvent[]) => Promise<void> | void
-  onBeforeTask?: (this: PageAgent) => Promise<void> | void
-  onAfterTask?: (this: PageAgent, result: ExecutionResult) => Promise<void> | void
-  onDispose?: (this: PageAgent, reason?: string) => void
-
-  /**
-   * Transform page content before sending to LLM.
-   * Use cases: inspect extraction results, modify page info, mask sensitive data.
-   */
-  transformPageContent?: (content: string) => Promise<string> | string
-
-  /** @experimental Enable JavaScript execution tool */
-  experimentalScriptExecutionTool?: boolean
-}`}
-				/>
-			</section>
-
-			{/* PageController Configuration */}
-			<section className="mb-10">
-				<h2 className="text-2xl font-semibold mb-4">
-					{isZh ? 'PageController 配置' : 'PageController Configuration'}
-				</h2>
-				<p className="text-gray-600 dark:text-gray-400 mb-4">
-					{isZh
-						? '配置 DOM 提取、元素交互和视觉高亮的细节。'
-						: 'Configure DOM extraction, element interaction, and visual highlighting.'}
-				</p>
-				<CodeEditor
-					className="mb-4"
-					language="typescript"
-					code={`interface DomConfig {
-  /** Elements to exclude from interaction */
-  interactiveBlacklist?: (Element | (() => Element))[]
-
-  /** Elements to force include for interaction */
-  interactiveWhitelist?: (Element | (() => Element))[]
-
-  /** Additional attributes to include in DOM extraction */
-  include_attributes?: string[]
-
-  /** Highlight overlay opacity (0-1) */
-  highlightOpacity?: number
-
-  /** Highlight label opacity (0-1) */
-  highlightLabelOpacity?: number
-}
-
-interface PageControllerConfig extends DomConfig {
-  /** Viewport expansion in pixels */
-  viewportExpansion?: number
-
-  /** Enable visual mask overlay during operations (default: false) */
-  enableMask?: boolean
-}`}
-				/>
-			</section>
-
-			{/* Complete Type */}
-			<section className="mb-10">
-				<h2 className="text-2xl font-semibold mb-4">{isZh ? '完整类型' : 'Complete Type'}</h2>
-				<CodeEditor
-					language="typescript"
-					code={`type PageAgentConfig = LLMConfig & AgentConfig & PageControllerConfig`}
-				/>
-			</section>
-
-			{/* Programmatic Usage Example */}
-			<section className="mb-10">
-				<h2 className="text-2xl font-semibold mb-4">
-					{isZh ? '程序化使用配置' : 'Programmatic Usage'}
-				</h2>
-				<p className="text-gray-600 dark:text-gray-400 mb-4">
-					{isZh
-						? '对于程序化集成场景，可以禁用 UI。'
-						: 'For programmatic integration, you can disable UI.'}
-				</p>
-				<CodeEditor
-					language="typescript"
-					code={`const agent = new PageAgent({
-  baseURL: 'https://api.openai.com/v1',
-  apiKey: 'your-api-key',
-  model: 'your-model-name',
-
-  // Disable all UI features for pure programmatic usage
-  enablePanel: false,        // Don't create Panel UI
-  enableMask: false,         // Don't show visual overlay (mask and pointer)
-  // enableAskUser is automatically disabled when enablePanel is false
-
-  // Or keep Panel but disable post-task prompts
-  // enablePanel: true,
-  // promptForNextTask: false,
-})
-
-// Pure programmatic execution
-const result = await agent.execute('search for TypeScript documentation')
-console.log(result.success, result.data, result.history)`}
-				/>
-			</section>
-		</div>
-	)
-}
--- a/packages/website/tsconfig.json
+++ b/packages/website/tsconfig.json
@@ -10,11 +10,12 @@
            // Self root
            "@/*": ["src/*"],

-            // Simplified monorepo solution (raw npm workspace with hoisting)
-            "page-agent": ["../page-agent/src/PageAgent.ts"],
            "@page-agent/llms": ["../llms/src/index.ts"],
            "@page-agent/page-controller": ["../page-controller/src/PageController.ts"],
-            "@page-agent/ui": ["../ui/src/index.ts"]
+            "@page-agent/core": ["../core/src/PageAgentCore.ts"],
+            "@page-agent/ui": ["../ui/src/index.ts"],
+
+            "page-agent": ["../page-agent/src/PageAgent.ts"]
        }
    },
    "include": ["**/*.ts", "**/*.tsx"],
@@ -22,8 +23,10 @@
    "references": [
        //
        { "path": "../llms" },
-        { "path": "../page-agent" },
        { "path": "../page-controller" },
-        { "path": "../ui" }
+        { "path": "../core" },
+        { "path": "../ui" },
+
+        { "path": "../page-agent" }
    ]
 }
--- a/packages/website/vite.config.js
+++ b/packages/website/vite.config.js
@@ -36,9 +36,11 @@ export default defineConfig(({ mode }) => ({
 			'@': resolve(__dirname, 'src'),

 			// Monorepo packages (always bundle local code instead of npm versions)
-			'@page-agent/llms': resolve(__dirname, '../llms/src/index.ts'),
 			'@page-agent/page-controller': resolve(__dirname, '../page-controller/src/PageController.ts'),
+			'@page-agent/llms': resolve(__dirname, '../llms/src/index.ts'),
+			'@page-agent/core': resolve(__dirname, '../core/src/PageAgentCore.ts'),
 			'@page-agent/ui': resolve(__dirname, '../ui/src/index.ts'),
+
 			'page-agent': resolve(__dirname, '../page-agent/src/PageAgent.ts'),
 		},
 	},