Agent/debug_cache.py

"""
调试脚本：检查缓存失效的原因

分析trace中的消息，看看为什么缓存点没有生效
"""
import json
from pathlib import Path
import hashlib

trace_dir = Path("/Users/elksmmx/Desktop/agent 2.10/Agent/examples/find knowledge/.trace/7ff963bc-3106-47fc-9725-b712f3e0d7d5/messages")

print("=== 分析缓存失效原因 ===\n")

# 读取前几次assistant消息，看看它们发送的消息列表是否一致
assistant_calls = []
for i in range(1, 101):
    fname = trace_dir / f"7ff963bc-3106-47fc-9725-b712f3e0d7d5-{i:04d}.json"
    if not fname.exists():
        continue

    with open(fname) as f:
        data = json.load(f)

    if data.get('role') == 'assistant':
        assistant_calls.append({
            'seq': i,
            'cache_read': data.get('cache_read_tokens', 0),
            'cache_create': data.get('cache_creation_tokens', 0),
        })

print("=== Assistant消息的缓存情况 ===")
for call in assistant_calls[:20]:  # 只看前20次
    seq = call['seq']
    cache_read = call['cache_read']
    cache_create = call['cache_create']

    if cache_create > 0:
        print(f"seq={seq:03d} 🆕 创建缓存 {cache_create}")
    elif cache_read > 0:
        print(f"seq={seq:03d} ✅ 命中缓存 {cache_read}")
    else:
        print(f"seq={seq:03d} ❌ 缓存失效")

print("\n=== 关键发现 ===")
print(f"总共{len(assistant_calls)}次assistant调用")
print(f"创建缓存: {sum(1 for c in assistant_calls if c['cache_create'] > 0)}次")
print(f"命中缓存: {sum(1 for c in assistant_calls if c['cache_read'] > 0)}次")
print(f"缓存失效: {sum(1 for c in assistant_calls if c['cache_read'] == 0 and c['cache_create'] == 0)}次")

# 分析缓存创建的token数
cache_creates = [c['cache_create'] for c in assistant_calls if c['cache_create'] > 0]
if cache_creates:
    print(f"\n缓存创建的token数: {cache_creates}")
    print(f"所有缓存大小都是 {cache_creates[0]} tokens，说明只缓存了system prompt")

# 分析缓存命中的token数
cache_reads = [c['cache_read'] for c in assistant_calls if c['cache_read'] > 0]
if cache_reads:
    unique_reads = set(cache_reads)
    print(f"\n缓存命中的token数: {unique_reads}")
    if len(unique_reads) == 1:
        print(f"所有缓存命中都是 {list(unique_reads)[0]} tokens，说明只命中了system prompt")

print("\n=== 结论 ===")
print("虽然我们在第20、40、60条消息上设置了缓存点，")
print("但Anthropic API只创建了system prompt的缓存。")
print("这说明后续消息的缓存点没有生效。")
print("\n可能的原因：")
print("1. 消息内容在每次调用时发生了变化")
print("2. 消息格式转换（str->list）导致Anthropic认为是不同的消息")
print("3. Anthropic的缓存机制有其他限制（如最小缓存长度）")