deepagents----/tests/analyze_llm_calls.py

"""
分析LLM调用记录

使用方法：
    python tests/analyze_llm_calls.py tests/llm_calls_20251031_150543.json
"""

import sys
import json


def analyze_llm_calls(json_file):
    """分析LLM调用记录"""
    with open(json_file, 'r', encoding='utf-8') as f:
        data = json.load(f)

    print("\n" + "="*80)
    print("LLM调用分析报告")
    print("="*80)

    print(f"\n总调用次数: {data['total_calls']}")

    for i, call in enumerate(data['calls'], 1):
        print(f"\n{'─'*80}")
        print(f"调用 #{i}")
        print('─'*80)

        # 时间信息
        start = call.get('timestamp_start', 'N/A')
        end = call.get('timestamp_end', 'N/A')
        print(f"时间: {start} -> {end}")

        # 消息数
        messages = call.get('messages', [[]])
        if messages:
            msg_count = len(messages[0])
            print(f"输入消息数: {msg_count}")

            # 显示最后一条消息类型
            if messages[0]:
                last_msg = messages[0][-1]
                print(f"最后一条输入消息: {last_msg['type']}")

        # 响应信息
        response = call.get('response', {})
        generations = response.get('generations', [])

        if generations:
            gen = generations[0]
            msg = gen.get('message', {})

            print(f"响应类型: {msg.get('type', 'N/A')}")

            # 内容
            content = msg.get('content', '')
            if content:
                preview = content[:100].replace('\n', ' ')
                print(f"响应内容: {preview}...")

            # 工具调用
            tool_calls = msg.get('tool_calls', [])
            if tool_calls:
                print(f"工具调用: {len(tool_calls)} 个")
                for tc in tool_calls:
                    print(f"  - {tc['name']}")
            else:
                print("工具调用: 无")

        # Token使用
        llm_output = response.get('llm_output', {})
        token_usage = llm_output.get('token_usage', {})
        if token_usage:
            print(f"Token使用: {token_usage.get('prompt_tokens', 0)} input + {token_usage.get('completion_tokens', 0)} output = {token_usage.get('total_tokens', 0)} total")

    print("\n" + "="*80)
    print("执行流程总结")
    print("="*80)

    # 分析执行流程
    call_summaries = []
    for i, call in enumerate(data['calls'], 1):
        response = call.get('response', {})
        generations = response.get('generations', [])

        if generations:
            msg = generations[0].get('message', {})
            tool_calls = msg.get('tool_calls', [])

            if tool_calls:
                tools = [tc['name'] for tc in tool_calls]
                call_summaries.append(f"调用#{i}: {', '.join(tools)}")
            else:
                content_preview = msg.get('content', '')[:50].replace('\n', ' ')
                call_summaries.append(f"调用#{i}: 返回文本 ({content_preview}...)")

    for summary in call_summaries:
        print(f"  {summary}")

    # 判断是否完成
    print("\n" + "="*80)
    print("状态判断")
    print("="*80)

    last_call = data['calls'][-1]
    last_response = last_call.get('response', {})
    last_generations = last_response.get('generations', [])

    if last_generations:
        last_msg = last_generations[0].get('message', {})
        last_tool_calls = last_msg.get('tool_calls', [])

        if not last_tool_calls:
            print("⚠️  最后一次调用没有工具调用")
            print("原因: SubAgent返回了纯文本响应，导致主Agent停止")
            print("影响: Agent停止执行，未完成完整流程")
            print("\n预期行为: 主Agent应该继续执行步骤3（并行搜索）")
        else:
            print("✅ 最后一次调用有工具调用，流程继续")
    else:
        print("❌ 无法判断状态")

    # 检查是否完成意图分析
    search_queries_created = False
    for call in data['calls']:
        response = call.get('response', {})
        generations = response.get('generations', [])
        if generations:
            msg = generations[0].get('message', {})
            tool_calls = msg.get('tool_calls', [])
            for tc in tool_calls:
                if tc['name'] == 'write_file' and '/search_queries.json' in str(tc.get('args', {})):
                    search_queries_created = True

    print("\n" + "="*80)
    print("步骤完成情况")
    print("="*80)
    print(f"✅ 步骤1: 初始化 - 已完成 (/question.txt, /config.json)")
    print(f"✅ 步骤2: 意图分析 - {'已完成' if search_queries_created else '未完成'} (/search_queries.json)")
    print(f"❌ 步骤3: 并行搜索 - 未开始")
    print(f"❌ 后续步骤 - 未开始")

    print("\n" + "="*80)
    print("建议")
    print("="*80)
    print("1. 问题根源: intent-analyzer SubAgent完成后返回纯文本，导致主Agent停止")
    print("2. 解决方案: 修改主Agent的系统提示词，明确要求在SubAgent返回后继续执行下一步")
    print("3. 或者: 检查LangGraph的recursion_limit配置，确保允许足够的步骤数")


if __name__ == "__main__":
    if len(sys.argv) < 2:
        print("使用方法: python analyze_llm_calls.py <json_file>")
        sys.exit(1)

    json_file = sys.argv[1]
    analyze_llm_calls(json_file)