# -*- coding: utf-8 -*-
"""对指定会话做「积分消耗」归因：
  ① 逐请求 input/output token 曲线（credit burn 的直接证据）
  ② 上下文构成（谁占的）
  ③ 最重的工具 / 重复读的文件
  ④ 用户可见输出 vs 烧掉的输入
用法: python diag-credit.py <sid前8位>
"""
import io
import json
import os
import re
import sys
from collections import Counter

PROJ = r"E:\ProgramData\.workbuddy\projects\e-ProgramData-AIProject-aliyun-dsh-server"
SID8 = sys.argv[1] if len(sys.argv) > 1 else "7057685c"
P = None
for fn in os.listdir(PROJ):
    if fn.startswith(SID8) and fn.endswith(".jsonl"):
        P = os.path.join(PROJ, fn)
assert P, "找不到转录 " + SID8
CPT = 2.5


def sz(x):
    if isinstance(x, str):
        return len(x)
    try:
        return len(json.dumps(x, ensure_ascii=False))
    except Exception:
        return len(str(x))


def txt_of(c):
    if isinstance(c, list):
        return "".join(x.get("text", "") for x in c if isinstance(x, dict))
    return str(c or "")


titles, usages, tot = [], [], Counter()
results, calls, reads = [], [], Counter()
per_tool_chars, per_tool_n = Counter(), Counter()
asst_msgs, user_real = [], []
pending_call = {}

with io.open(P, encoding="utf-8", errors="replace") as f:
    for ln in f:
        if not ln.strip():
            continue
        try:
            o = json.loads(ln)
        except Exception:
            continue
        t = o.get("type")
        if t == "ai-title":
            titles.append(o.get("aiTitle", ""))
        elif t == "message":
            s = sz(txt_of(o.get("content")))
            if o.get("role") == "user":
                tot["user_msg"] += s
                txt = txt_of(o.get("content"))
                m = re.search(r"<user_query>(.*?)</user_query>", txt, re.S)
                if m:
                    user_real.append((o.get("timestamp"), " ".join(m.group(1).split())[:100]))
            else:
                tot["asst_msg"] += s
                asst_msgs.append((o.get("timestamp"), sz(txt_of(o.get("content")))))
        elif t == "reasoning":
            tot["reasoning"] += sz(o.get("content"))
        elif t == "function_call":
            s = sz(o.get("arguments"))
            tot["fn_call_args"] += s
            nm = o.get("name") or "?"
            calls.append((nm, s))
            per_tool_chars[nm] += s
            per_tool_n[nm] += 1
            u = ((o.get("message") or {}).get("usage")) or {}
            if u:
                usages.append((o.get("timestamp"), u.get("input_tokens"), u.get("output_tokens"),
                               u.get("cache_read_input_tokens")))
            if nm == "Read":
                try:
                    a = json.loads(o.get("arguments") or "{}")
                    fp = a.get("file_path") or a.get("path") or ""
                    if fp:
                        reads[os.path.basename(fp)] += 1
                except Exception:
                    pass
        elif t == "function_call_result":
            s = sz(o.get("output"))
            tot["fn_result"] += s
            nm = o.get("name") or "?"
            per_tool_chars[nm] += s
            per_tool_n[nm] += 1
            results.append((nm, s, " ".join(str(o.get("output"))[:100].split())))
        else:
            tot["other:" + str(t)] += sz(o)

print("=" * 78)
print("会话 %s  标题: %s" % (SID8, titles))
print("=" * 78)

print()
print("① 逐请求 token（input / output / 缓存读）—— 直接看积分去向")
print("=" * 78)
if usages:
    ti = sum(u[1] or 0 for u in usages)
    to = sum(u[2] or 0 for u in usages)
    tc = sum(u[3] or 0 for u in usages)
    print("  请求数 = %d" % len(usages))
    print("  Σ input  = %12d token" % ti)
    print("  Σ output = %12d token" % to)
    print("  Σ cache_read = %12d token（%.0f%% 的 input 命中缓存）" % (tc, 100.0 * tc / max(ti, 1)))
    print("  ★ input : output = %.1f : 1" % (ti / max(to, 1)))
    print()
    print("  前 5 条 / 后 5 条：")
    for u in usages[:5] + [None] + usages[-5:]:
        if u is None:
            print("    …")
        else:
            print("    ts=%-14s in=%-8s out=%-6s cache=%s" % (str(u[0])[:14], u[1], u[2], u[3]))
else:
    print("  （无 usage）")

print()
print("② 上下文构成（转录体量归因）")
print("=" * 78)
grand = sum(tot.values())
for k, v in tot.most_common():
    print("  %-30s %10d 字符 ≈%9d tok  %5.1f%%" % (k, v, int(v / CPT), 100.0 * v / grand))
print()
print("  AI 正文 : 用户真实输入 比 = %.1f : 1" % (tot["asst_msg"] / max(tot["user_msg"], 1)))
print("  工具往返占体量 = %.1f%%（fn_result + fn_call_args）"
      % (100.0 * (tot["fn_result"] + tot["fn_call_args"]) / max(grand, 1)))

print()
print("③ 最重的工具（累计字符 / 次数）")
print("=" * 78)
for nm, c in per_tool_chars.most_common(12):
    print("  %-18s %10d 字符 ≈%9d tok   调用 %d 次  (平均 %d 字符)"
          % (nm, c, int(c / CPT), per_tool_n[nm], c // max(per_tool_n[nm], 1)))

print()
print("④ 单条最大工具结果 Top 10")
print("=" * 78)
for i, (nm, c, pv) in enumerate(sorted(results, key=lambda x: -x[1])[:10], 1):
    print("  %2d) %-14s %8d 字符 ≈%6d tok | %s" % (i, nm, c, int(c / CPT), pv[:86]))

print()
print("⑤ 被重复 Read 的文件（读一次留一份，永不释放）")
print("=" * 78)
rep = [(k, v) for k, v in reads.most_common(12) if v >= 2]
for k, v in rep:
    print("  %-52s %d 次" % (k[:52], v))
if not rep:
    print("  （无重复读）")

print()
print("⑥ 用户真实输入（<user_query>）共 %d 条" % len(user_real))
print("=" * 78)
for ts, q in user_real[:30]:
    print("  [%s] %s" % (ts, q))