guantao 1 неделя назад
Родитель
Сommit
7c9fa40add

BIN
examples/deep_research/input/1.jpeg


+ 55 - 0
examples/deep_research/input/1_invariant_features.json

@@ -0,0 +1,55 @@
+{
+  "主色调Hex序列": [
+    {
+      "hex": "#3D5C3D",
+      "占比": "25.3%"
+    },
+    {
+      "hex": "#6B8E6B",
+      "占比": "20.1%"
+    },
+    {
+      "hex": "#FFFFFF",
+      "占比": "15.7%"
+    },
+    {
+      "hex": "#A0C0A0",
+      "占比": "12.5%"
+    },
+    {
+      "hex": "#80A080",
+      "占比": "9.8%"
+    }
+  ],
+  "色温估算值": 6000,
+  "高光RGB偏移量": [
+    0,
+    0,
+    0
+  ],
+  "中间调RGB偏移量": [
+    0,
+    0,
+    0
+  ],
+  "阴影RGB偏移量": [
+    0,
+    0,
+    0
+  ],
+  "饱和度量化值": 150,
+  "对比度量化值": 130,
+  "亮度量化值": 180,
+  "暗角强度": 0.0,
+  "颗粒噪声强度": 0.0,
+  "shadows色相偏移角度": 0,
+  "midtones色相偏移角度": 0,
+  "highlights色相偏移角度": 0,
+  "锐化程度": 0.5,
+  "胶片感特征": {
+    "褪色": false,
+    "漏光": false,
+    "色差": false
+  },
+  "LUT风格分类": "自然光照"
+}

BIN
examples/deep_research/input/3.jpeg


+ 55 - 0
examples/deep_research/input/3_invariant_features.json

@@ -0,0 +1,55 @@
+{
+  "主色调Hex序列": [
+    {
+      "color": "#4A6B2F",
+      "占比": "25.3%"
+    },
+    {
+      "color": "#8B9E60",
+      "占比": "20.1%"
+    },
+    {
+      "color": "#FFFFFF",
+      "占比": "15.7%"
+    },
+    {
+      "color": "#D4D4D4",
+      "占比": "10.5%"
+    },
+    {
+      "color": "#A07E4C",
+      "占比": "8.2%"
+    }
+  ],
+  "色温估算值": 6500,
+  "高光RGB偏移量": [
+    0,
+    0,
+    0
+  ],
+  "中间调RGB偏移量": [
+    0,
+    0,
+    0
+  ],
+  "阴影RGB偏移量": [
+    0,
+    0,
+    0
+  ],
+  "饱和度量化值": 150,
+  "对比度量化值": 130,
+  "亮度量化值": 180,
+  "暗角强度": 0.1,
+  "颗粒噪声强度": 0.05,
+  "shadows色相偏移角度": 0,
+  "midtones色相偏移角度": 0,
+  "highlights色相偏移角度": 0,
+  "锐化程度": 0.3,
+  "胶片感特征": {
+    "褪色": false,
+    "漏光": false,
+    "色差": false
+  },
+  "LUT风格分类": "Natural"
+}

BIN
examples/deep_research/input/7.jpeg


+ 59 - 0
examples/deep_research/input/7_invariant_features.json

@@ -0,0 +1,59 @@
+{
+  "dominant_colors": [
+    {
+      "hex": "#3E6B3E",
+      "percentage": 25.3
+    },
+    {
+      "hex": "#FFFFFF",
+      "percentage": 20.1
+    },
+    {
+      "hex": "#8C8C8C",
+      "percentage": 15.7
+    },
+    {
+      "hex": "#D4D4D4",
+      "percentage": 12.5
+    },
+    {
+      "hex": "#A0A0A0",
+      "percentage": 9.8
+    }
+  ],
+  "color_temperature_kelvin": 6000,
+  "rgb_offset": {
+    "highlights": {
+      "R": 5,
+      "G": 0,
+      "B": -5
+    },
+    "midtones": {
+      "R": 0,
+      "G": 0,
+      "B": 0
+    },
+    "shadows": {
+      "R": -5,
+      "G": 0,
+      "B": 5
+    }
+  },
+  "saturation": 150,
+  "contrast": 130,
+  "brightness": 180,
+  "vignette_intensity": 0.05,
+  "grain_intensity": 0.0,
+  "hue_shift": {
+    "shadows": 0,
+    "midtones": 0,
+    "highlights": 0
+  },
+  "sharpening_degree": 0.6,
+  "film_characteristics": {
+    "fade": false,
+    "light_leak": false,
+    "chromatic_aberration": false
+  },
+  "lut_style_classification": "Natural_Soft"
+}

+ 9 - 0
examples/deep_research/input/《秋日际遇》写生油画.json

@@ -0,0 +1,9 @@
+{
+  "images": [
+    "examples/deep_research/input/1.jpeg",
+    "examples/deep_research/input/3.jpeg",
+    "examples/deep_research/input/7.jpeg"
+  ],
+  "body_text": "听闻秋日是倒放的春天\n于是我心中有一座秋日的花园\n栽种着一簇簇淡却温暖的花\n风沿着远边的山吹来\n热情的阳光里秋风微凉\n与颜料一起酝酿出的画面\n白裙是一抹无暇\n迎着光绘画出\n那片在我心上开满\n限定的浪漫\n被画架支起\n绿草坪还驻留了匆匆而过的热闹\n再添一笔白\n为我画一枝玫瑰的奇遇\n———@万淮 #草地拍照[话题]##画画[话题]#",
+  "title": "《秋日际遇》写生油画"
+}

+ 23 - 0
examples/deep_research/production.prompt

@@ -0,0 +1,23 @@
+---
+model: sonnet-4.6
+temperature: 0.3
+---
+
+$system$
+你是一个顶尖的多模态特征工程专家。你的核心任务是利用技术手段(而非自然语言)从原始素材中提取出可复用的“视觉/听觉灵魂”,并验证这些特征是否能指导生成模型还原出一致的内容。
+你的行动准则:
+拒绝平庸描述:严禁使用“构图精美”、“色彩柔和”、“动作自然”等文学化词汇。你必须提供硬核数据表示,如:主色调 Hex 序列、关键点坐标矩阵 $(x, y)$、边缘拓扑图、或者是显著性热力图。
+工具驱动思维:在接到任务后,首先调研并调用最适合的工具链(如:OpenCV 图像处理、MediaPipe 姿态检测、DINOv2 特征提取、Nano Banana 图像重组)。
+闭环还原验证:所有的特征提取必须通过“还原测试”。如果你提取的特征(如人物骨架)能让生成模型在不同背景下复刻出相同的动作,则视为成功。
+工程化交付:你不仅要给出结论,还要生成可执行的 Python 脚本(resource 目录下)和结构化的数据文件(feature.json)。
+$user$
+**任务目标**
+分析 examples/deep_research/input/ 中的原内容,针对特征维度 滤镜风格【在此处输入维度,例如:滤镜风格、人物骨架拓扑、画面几何构图】 进行深度提取与还原验证。
+**交付清单** (Output Requirement)
+在 examples/deep_research/output_1/ 目录下输出:
+    1. method.md:记录提取该维度所用的具体算法工具(如 OpenCV Canny/MediaPipe)及还原效果量化评估。
+    2. feature.json:存储提取出的结构化特征数据(如归一化的坐标、颜色值等),禁止自然语言。
+    3. reduction.jpg:采用三段式对比展示:[ 原图 | 提取的特征图(或点位可视化) | 还原后的生成图 ]。
+    4. ./resource/:保存实现上述提取过程的简单 Python 脚本或配置文件。
+禁止降级解决,请立即开始调研并调用工具执行。
+不要用sandbox执行,直接在本地目录执行(examples/deep_research)

+ 581 - 0
examples/deep_research/run.py

@@ -0,0 +1,581 @@
+"""
+示例(增强版)
+
+使用 Agent 模式 + Skills
+
+新增功能:
+1. 支持命令行随时打断(输入 'p' 暂停,'q' 退出)
+2. 暂停后可插入干预消息
+3. 支持触发经验总结
+4. 查看当前 GoalTree
+5. 框架层自动清理不完整的工具调用
+6. 支持通过 --trace <ID> 恢复已有 Trace 继续执行
+"""
+
+import argparse
+import os
+import sys
+import select
+import asyncio
+from pathlib import Path
+
+# Clash Verge TUN 模式兼容:禁止 httpx/urllib 自动检测系统 HTTP 代理
+# TUN 虚拟网卡已在网络层接管所有流量,不需要应用层再走 HTTP 代理,
+# 否则 httpx 检测到 macOS 系统代理 (127.0.0.1:7897) 会导致 ConnectError
+os.environ.setdefault("no_proxy", "*")
+
+# 添加项目根目录到 Python 路径
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+from dotenv import load_dotenv
+load_dotenv()
+
+from agent.llm.prompts import SimplePrompt
+from agent.core.runner import AgentRunner, RunConfig
+from agent.core.presets import AgentPreset, register_preset
+from agent.trace import (
+    FileSystemTraceStore,
+    Trace,
+    Message,
+)
+from agent.llm import create_openrouter_llm_call
+from agent.tools import get_tool_registry
+
+
+# ===== 非阻塞 stdin 检测 =====
+if sys.platform == 'win32':
+    import msvcrt
+
+def check_stdin() -> str | None:
+    """
+    跨平台非阻塞检查 stdin 输入。
+    Windows: 使用 msvcrt.kbhit()
+    macOS/Linux: 使用 select.select()
+    """
+    if sys.platform == 'win32':
+        # 检查是否有按键按下
+        if msvcrt.kbhit():
+            # 读取按下的字符(msvcrt.getwch 是非阻塞读取宽字符)
+            ch = msvcrt.getwch().lower()
+            if ch == 'p':
+                return 'pause'
+            if ch == 'q':
+                return 'quit'
+            # 如果是其他按键,可以选择消耗掉或者忽略
+        return None
+    else:
+        # Unix/Mac 逻辑
+        ready, _, _ = select.select([sys.stdin], [], [], 0)
+        if ready:
+            line = sys.stdin.readline().strip().lower()
+            if line in ('p', 'pause'):
+                return 'pause'
+            if line in ('q', 'quit'):
+                return 'quit'
+        return None
+
+
+# ===== 交互菜单 =====
+
+def _read_multiline() -> str:
+    """
+    读取多行输入,以连续两次回车(空行)结束。
+
+    单次回车只是换行,不会提前终止输入。
+    """
+    print("\n请输入干预消息(连续输入两次回车结束):")
+    lines: list[str] = []
+    blank_count = 0
+    while True:
+        line = input()
+        if line == "":
+            blank_count += 1
+            if blank_count >= 2:
+                break
+            lines.append("")          # 保留单个空行
+        else:
+            blank_count = 0
+            lines.append(line)
+
+    # 去掉尾部多余空行
+    while lines and lines[-1] == "":
+        lines.pop()
+    return "\n".join(lines)
+
+
+async def show_interactive_menu(
+    runner: AgentRunner,
+    trace_id: str,
+    current_sequence: int,
+    store: FileSystemTraceStore,
+):
+    """
+    显示交互式菜单,让用户选择操作。
+
+    进入本函数前不再有后台线程占用 stdin,所以 input() 能正常工作。
+    """
+    print("\n" + "=" * 60)
+    print("  执行已暂停")
+    print("=" * 60)
+    print("请选择操作:")
+    print("  1. 插入干预消息并继续")
+    print("  2. 触发经验总结(reflect)")
+    print("  3. 查看当前 GoalTree")
+    print("  4. 手动压缩上下文(compact)")
+    print("  5. 继续执行")
+    print("  6. 停止执行")
+    print("=" * 60)
+
+    while True:
+        choice = input("请输入选项 (1-6): ").strip()
+
+        if choice == "1":
+            text = _read_multiline()
+            if not text:
+                print("未输入任何内容,取消操作")
+                continue
+
+            print(f"\n将插入干预消息并继续执行...")
+            # 从 store 读取实际的 last_sequence,避免本地 current_sequence 过时
+            live_trace = await store.get_trace(trace_id)
+            actual_sequence = live_trace.last_sequence if live_trace and live_trace.last_sequence else current_sequence
+            return {
+                "action": "continue",
+                "messages": [{"role": "user", "content": text}],
+                "after_sequence": actual_sequence,
+            }
+
+        elif choice == "2":
+            # 触发经验总结
+            print("\n触发经验总结...")
+            focus = input("请输入反思重点(可选,直接回车跳过): ").strip()
+
+            from agent.trace.compaction import build_reflect_prompt
+
+            # 保存当前 head_sequence
+            trace = await store.get_trace(trace_id)
+            saved_head = trace.head_sequence
+
+            prompt = build_reflect_prompt()
+            if focus:
+                prompt += f"\n\n请特别关注:{focus}"
+
+            print("正在生成反思...")
+            reflect_cfg = RunConfig(trace_id=trace_id, max_iterations=1, tools=[])
+
+            reflection_text = ""
+            try:
+                result = await runner.run_result(
+                    messages=[{"role": "user", "content": prompt}],
+                    config=reflect_cfg,
+                )
+                reflection_text = result.get("summary", "")
+            finally:
+                # 恢复 head_sequence(反思消息成为侧枝)
+                await store.update_trace(trace_id, head_sequence=saved_head)
+
+            # 追加到 experiences 文件
+            if reflection_text:
+                from datetime import datetime
+                experiences_path = runner.experiences_path or "./.cache/experiences_find.md"
+                os.makedirs(os.path.dirname(experiences_path), exist_ok=True)
+                header = f"\n\n---\n\n## {trace_id} ({datetime.now().strftime('%Y-%m-%d %H:%M')})\n\n"
+                with open(experiences_path, "a", encoding="utf-8") as f:
+                    f.write(header + reflection_text + "\n")
+                print(f"\n反思已保存到: {experiences_path}")
+                print("\n--- 反思内容 ---")
+                print(reflection_text)
+                print("--- 结束 ---\n")
+            else:
+                print("未生成反思内容")
+
+            continue
+
+        elif choice == "3":
+            goal_tree = await store.get_goal_tree(trace_id)
+            if goal_tree and goal_tree.goals:
+                print("\n当前 GoalTree:")
+                print(goal_tree.to_prompt())
+            else:
+                print("\n当前没有 Goal")
+            continue
+
+        elif choice == "4":
+            # 手动压缩上下文
+            print("\n正在执行上下文压缩(compact)...")
+            try:
+                goal_tree = await store.get_goal_tree(trace_id)
+                trace = await store.get_trace(trace_id)
+                if not trace:
+                    print("未找到 Trace,无法压缩")
+                    continue
+
+                # 重建当前 history
+                main_path = await store.get_main_path_messages(trace_id, trace.head_sequence)
+                history = [msg.to_llm_dict() for msg in main_path]
+                head_seq = main_path[-1].sequence if main_path else 0
+                next_seq = head_seq + 1
+
+                compact_config = RunConfig(trace_id=trace_id)
+                new_history, new_head, new_seq = await runner._compress_history(
+                    trace_id=trace_id,
+                    history=history,
+                    goal_tree=goal_tree,
+                    config=compact_config,
+                    sequence=next_seq,
+                    head_seq=head_seq,
+                )
+                print(f"\n✅ 压缩完成: {len(history)} 条消息 → {len(new_history)} 条")
+            except Exception as e:
+                print(f"\n❌ 压缩失败: {e}")
+            continue
+
+        elif choice == "5":
+            print("\n继续执行...")
+            return {"action": "continue"}
+
+        elif choice == "6":
+            print("\n停止执行...")
+            return {"action": "stop"}
+
+        else:
+            print("无效选项,请重新输入")
+
+
+async def main():
+    # 解析命令行参数
+    parser = argparse.ArgumentParser(description="任务 (Agent 模式 + 交互增强)")
+    parser.add_argument(
+        "--trace", type=str, default=None,
+        help="已有的 Trace ID,用于恢复继续执行(不指定则新建)",
+    )
+    args = parser.parse_args()
+
+    # 路径配置
+    base_dir = Path(__file__).parent
+    project_root = base_dir.parent.parent
+    prompt_path = base_dir / "production.prompt"
+
+    # 为每次运行创建独立的输出目录(基于时间戳或 trace_id)
+    if args.trace:
+        # 恢复模式:使用已有 trace_id 作为目录名
+        output_dir = base_dir / "outputs" / args.trace[:8]
+    else:
+        # 新建模式:使用时间戳
+        from datetime import datetime
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        output_dir = base_dir / "outputs" / timestamp
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    print(f"📁 输出目录: {output_dir}")
+    print()
+
+    # 加载项目级 presets(examples/how/presets.json)
+    presets_path = base_dir / "presets.json"
+    if presets_path.exists():
+        import json
+        with open(presets_path, "r", encoding="utf-8") as f:
+            project_presets = json.load(f)
+        for name, cfg in project_presets.items():
+            register_preset(name, AgentPreset(**cfg))
+        print(f"   - 已加载项目 presets: {list(project_presets.keys())}")
+
+    # Skills 目录(可选:用户自定义 skills)
+    # 注意:内置 skills(agent/memory/skills/)会自动加载
+    skills_dir = str(base_dir / "skills")
+
+    print("=" * 60)
+    print("mcp/skills 发现、获取、评价 分析任务 (Agent 模式 + 交互增强)")
+    print("=" * 60)
+    print()
+    print("💡 交互提示:")
+    print("   - 执行过程中输入 'p' 或 'pause' 暂停并进入交互模式")
+    print("   - 执行过程中输入 'q' 或 'quit' 停止执行")
+    print("=" * 60)
+    print()
+
+    # 1. 加载 prompt
+    print("1. 加载 prompt 配置...")
+    prompt = SimplePrompt(prompt_path)
+
+    # 2. 构建消息(仅新建时使用,恢复时消息已在 trace 中)
+    print("2. 构建任务消息...")
+    # 注入输出目录到 prompt 变量
+    messages = prompt.build_messages(variables={
+        "output_dir": str(output_dir.relative_to(base_dir))
+    })
+
+    # 3. 创建 Agent Runner(配置 skills)
+    print("3. 创建 Agent Runner...")
+    print(f"   - Skills 目录: {skills_dir}")
+    print(f"   - 模型: {prompt.config.get('model', 'sonnet-4.5')}")
+
+    # 加载自定义工具
+    print("   - 加载自定义工具: nanobanana")
+    import examples.how.tool  # 导入自定义工具模块,触发 @tool 装饰器注册
+
+    store = FileSystemTraceStore(base_path=".trace")
+    runner = AgentRunner(
+        trace_store=store,
+        llm_call=create_openrouter_llm_call(model=f"anthropic/claude-{prompt.config.get('model', 'sonnet-4.5')}"),
+        skills_dir=skills_dir,
+        experiences_path="./.cache/experiences_find.md",
+        debug=True
+    )
+
+    # 4. 判断是新建还是恢复
+    resume_trace_id = args.trace
+    if resume_trace_id:
+        # 验证 trace 存在
+        existing_trace = await store.get_trace(resume_trace_id)
+        if not existing_trace:
+            print(f"\n错误: Trace 不存在: {resume_trace_id}")
+            sys.exit(1)
+        print(f"4. 恢复已有 Trace: {resume_trace_id[:8]}...")
+        print(f"   - 状态: {existing_trace.status}")
+        print(f"   - 消息数: {existing_trace.total_messages}")
+        print(f"   - 任务: {existing_trace.task}")
+    else:
+        print(f"4. 启动新 Agent 模式...")
+
+    print()
+
+    final_response = ""
+    current_trace_id = resume_trace_id
+    current_sequence = 0
+    should_exit = False
+
+    try:
+        # 恢复模式:不发送初始消息,只指定 trace_id 续跑
+        if resume_trace_id:
+            initial_messages = None  # None = 未设置,触发早期菜单检查
+            config = RunConfig(
+                model=f"claude-{prompt.config.get('model', 'sonnet-4.5')}",
+                temperature=float(prompt.config.get('temperature', 0.3)),
+                max_iterations=1000,
+                trace_id=resume_trace_id,
+            )
+        else:
+            initial_messages = messages
+            config = RunConfig(
+                model=f"claude-{prompt.config.get('model', 'sonnet-4.5')}",
+                temperature=float(prompt.config.get('temperature', 0.3)),
+                max_iterations=1000,
+                name="社交媒体内容解构、建构、评估任务",
+            )
+
+        while not should_exit:
+            # 如果是续跑,需要指定 trace_id
+            if current_trace_id:
+                config.trace_id = current_trace_id
+
+            # 清理上一轮的响应,避免失败后显示旧内容
+            final_response = ""
+
+            # 如果 trace 已完成/失败且没有新消息,直接进入交互菜单
+            # 注意:initial_messages 为 None 表示未设置(首次加载),[] 表示有意为空(用户选择"继续")
+            if current_trace_id and initial_messages is None:
+                check_trace = await store.get_trace(current_trace_id)
+                if check_trace and check_trace.status in ("completed", "failed"):
+                    if check_trace.status == "completed":
+                        print(f"\n[Trace] ✅ 已完成")
+                        print(f"  - Total messages: {check_trace.total_messages}")
+                        print(f"  - Total cost: ${check_trace.total_cost:.4f}")
+                    else:
+                        print(f"\n[Trace] ❌ 已失败: {check_trace.error_message}")
+                    current_sequence = check_trace.head_sequence
+
+                    menu_result = await show_interactive_menu(
+                        runner, current_trace_id, current_sequence, store
+                    )
+
+                    if menu_result["action"] == "stop":
+                        break
+                    elif menu_result["action"] == "continue":
+                        new_messages = menu_result.get("messages", [])
+                        if new_messages:
+                            initial_messages = new_messages
+                            config.after_sequence = menu_result.get("after_sequence")
+                        else:
+                            # 无新消息:对 failed trace 意味着重试,对 completed 意味着继续
+                            initial_messages = []
+                            config.after_sequence = None
+                        continue
+                    break
+
+                # 对 stopped/running 等非终态的 trace,直接续跑
+                initial_messages = []
+
+            print(f"{'▶️ 开始执行...' if not current_trace_id else '▶️ 继续执行...'}")
+
+            # 执行 Agent
+            paused = False
+            try:
+                async for item in runner.run(messages=initial_messages, config=config):
+                    # 检查用户中断
+                    cmd = check_stdin()
+                    if cmd == 'pause':
+                        # 暂停执行
+                        print("\n⏸️ 正在暂停执行...")
+                        if current_trace_id:
+                            await runner.stop(current_trace_id)
+
+                        # 等待一小段时间让 runner 处理 stop 信号
+                        await asyncio.sleep(0.5)
+
+                        # 显示交互菜单
+                        menu_result = await show_interactive_menu(
+                            runner, current_trace_id, current_sequence, store
+                        )
+
+                        if menu_result["action"] == "stop":
+                            should_exit = True
+                            paused = True
+                            break
+                        elif menu_result["action"] == "continue":
+                            # 检查是否有新消息需要插入
+                            new_messages = menu_result.get("messages", [])
+                            if new_messages:
+                                # 有干预消息,需要重新启动循环
+                                initial_messages = new_messages
+                                after_seq = menu_result.get("after_sequence")
+                                if after_seq is not None:
+                                    config.after_sequence = after_seq
+                                paused = True
+                                break
+                            else:
+                                # 没有新消息,需要重启执行
+                                initial_messages = []
+                                config.after_sequence = None
+                                paused = True
+                                break
+
+                    elif cmd == 'quit':
+                        print("\n🛑 用户请求停止...")
+                        if current_trace_id:
+                            await runner.stop(current_trace_id)
+                        should_exit = True
+                        break
+
+                    # 处理 Trace 对象(整体状态变化)
+                    if isinstance(item, Trace):
+                        current_trace_id = item.trace_id
+                        if item.status == "running":
+                            print(f"[Trace] 开始: {item.trace_id[:8]}...")
+                        elif item.status == "completed":
+                            print(f"\n[Trace] ✅ 完成")
+                            print(f"  - Total messages: {item.total_messages}")
+                            print(f"  - Total tokens: {item.total_tokens}")
+                            print(f"  - Total cost: ${item.total_cost:.4f}")
+                        elif item.status == "failed":
+                            print(f"\n[Trace] ❌ 失败: {item.error_message}")
+                        elif item.status == "stopped":
+                            print(f"\n[Trace] ⏸️ 已停止")
+
+                    # 处理 Message 对象(执行过程)
+                    elif isinstance(item, Message):
+                        current_sequence = item.sequence
+
+                        if item.role == "assistant":
+                            content = item.content
+                            if isinstance(content, dict):
+                                text = content.get("text", "")
+                                tool_calls = content.get("tool_calls")
+
+                                if text and not tool_calls:
+                                    # 纯文本回复(最终响应)
+                                    final_response = text
+                                    print(f"\n[Response] Agent 回复:")
+                                    print(text)
+                                elif text:
+                                    preview = text[:150] + "..." if len(text) > 150 else text
+                                    print(f"[Assistant] {preview}")
+
+                                if tool_calls:
+                                    for tc in tool_calls:
+                                        tool_name = tc.get("function", {}).get("name", "unknown")
+                                        print(f"[Tool Call] 🛠️  {tool_name}")
+
+                        elif item.role == "tool":
+                            content = item.content
+                            if isinstance(content, dict):
+                                tool_name = content.get("tool_name", "unknown")
+                                print(f"[Tool Result] ✅ {tool_name}")
+                            if item.description:
+                                desc = item.description[:80] if len(item.description) > 80 else item.description
+                                print(f"  {desc}...")
+
+            except Exception as e:
+                print(f"\n执行出错: {e}")
+                import traceback
+                traceback.print_exc()
+
+            # paused → 菜单已在暂停时内联显示过
+            if paused:
+                if should_exit:
+                    break
+                continue
+
+            # quit → 直接退出
+            if should_exit:
+                break
+
+            # Runner 退出(完成/失败/停止/异常)→ 显示交互菜单
+            if current_trace_id:
+                menu_result = await show_interactive_menu(
+                    runner, current_trace_id, current_sequence, store
+                )
+
+                if menu_result["action"] == "stop":
+                    break
+                elif menu_result["action"] == "continue":
+                    new_messages = menu_result.get("messages", [])
+                    if new_messages:
+                        initial_messages = new_messages
+                        config.after_sequence = menu_result.get("after_sequence")
+                    else:
+                        initial_messages = []
+                        config.after_sequence = None
+                    continue
+            break
+
+    except KeyboardInterrupt:
+        print("\n\n用户中断 (Ctrl+C)")
+        if current_trace_id:
+            await runner.stop(current_trace_id)
+
+    # 6. 输出结果
+    if final_response:
+        print()
+        print("=" * 60)
+        print("Agent 响应:")
+        print("=" * 60)
+        print(final_response)
+        print("=" * 60)
+        print()
+
+        # 7. 保存结果
+        output_file = output_dir / "result.txt"
+        with open(output_file, 'w', encoding='utf-8') as f:
+            f.write(final_response)
+
+        print(f"✓ 结果已保存到: {output_file}")
+        print()
+
+    # 可视化提示
+    if current_trace_id:
+        print("=" * 60)
+        print("可视化 Step Tree:")
+        print("=" * 60)
+        print("1. 启动 API Server:")
+        print("   python3 api_server.py")
+        print()
+        print("2. 浏览器访问:")
+        print("   http://localhost:8000/api/traces")
+        print()
+        print(f"3. Trace ID: {current_trace_id}")
+        print("=" * 60)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())