"""
=============================================================================
第 16 讲 · 代码 05（可选）：让真实的 Claude 当 GUM 的"提出"模块——从观察里推断用户
=============================================================================

对应讲义：lectures/lecture16_proactive_agents.md §7

【这个脚本要回答的问题】
    代码 01 里"把观察读成命题"是模拟的。真实的模型读同一份日志：
        1. 它会提出哪些关于用户的命题？置信度给得准不准（和隐藏画像比，算 Brier 分数）？
        2. 它会不会把敏感的事（求职、看心理医生）也写进去？让它按情境完整性自己标出敏感的命题，它标得对吗？
    一次调用：把虚构用户 Mira 一周的观察（文字）给它，要它返回最多 10 条命题（JSON）。

【需要什么】pip install anthropic；设置 ANTHROPIC_API_KEY（或 `ant auth login`）。**会花钱**：1 次调用。
    没有 SDK 或凭证时打印说明后正常退出。所有人、事都是虚构的。
【运行】
    python3 05_real_llm.py           # 默认 claude-opus-5-5，effort=low
    python3 05_real_llm.py --toy     # 不联网：用 proactive_lab 的模拟用户模型（检查脚本本身）
"""

import argparse
import json
import os
import re
import sys

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.join(HERE, "..", "lecture01"))
from console import print_header, print_observations, print_table  # noqa: E402
from proactive_lab import ALL, WEEKDAYS, build, make_log, words  # noqa: E402
from real_llm import DEFAULT_MODEL, ClaudeLM, sdk_available  # noqa: E402

SYSTEM = ("You build a user model from computer-use observations. Return JSON only: a list of up to 10 objects "
          '{"proposition": "...", "confidence": 0.0-1.0, "sensitive": true/false}. '
          "A proposition is a short claim about the user (habits, goals, preferences). Confidence should be calibrated: "
          "0.9 means you expect 9 in 10 such claims to be true. Mark sensitive=true for health, job-search, finances or "
          "other information the user would not want surfaced in a work context.")


def match(text):
    """把一条命题对应到隐藏画像里最像的那条（词重叠 ≥ 0.25），对应不上返回 None。"""
    w = words(text)
    best, bs = None, 0.25
    for k, f in ALL.items():
        sim = len(w & words(f.text)) / max(1, len(w | words(f.text)))
        if sim >= bs:
            best, bs = k, sim
    return best


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model", default=DEFAULT_MODEL)
    ap.add_argument("--effort", default="low")
    ap.add_argument("--toy", action="store_true")
    args = ap.parse_args()
    log = make_log("real", 7)
    lines = [f"Day {o.day + 1} ({WEEKDAYS[o.weekday]}) {o.hour}:00 — {o.text()}" for o in log]

    if args.toy:
        m = build(log, seed=0, audit=False)
        props = [{"proposition": q.text, "confidence": round(q.confidence, 2), "sensitive": ALL[q.key].kind == "sensitive"}
                 for q in m.top(10)]
        name = "toy（proactive_lab 的模拟用户模型，没有审计）"
    else:
        ok, why = sdk_available()
        if not ok:
            print(why)
            return
        try:
            lm = ClaudeLM(model=args.model, effort=args.effort, max_tokens=4000, stop_sequences=())
            out = lm([{"role": "system", "content": SYSTEM}, {"role": "user", "content": "Observations:\n" + "\n".join(lines)}])
        except Exception as e:  # noqa: BLE001
            if type(e).__name__ in ("AuthenticationError", "APIConnectionError"):
                print("凭证无效 / 连不上 API：设置 ANTHROPIC_API_KEY 或运行 ant auth login，并检查网络")
                return
            raise
        m = re.search(r"\[.*\]", out, flags=re.S)
        props = json.loads(m.group(0)) if m else []
        name = lm.name

    rows, sq, n, sens_ok, sens_n = [], 0.0, 0, 0, 0
    for p in props:
        k = match(p.get("proposition", ""))
        truth = ALL[k].true if k else None
        c = float(p.get("confidence", 0.5))
        if truth is not None:
            sq += (c - (1.0 if truth else 0.0)) ** 2
            n += 1
        is_sens = bool(k) and ALL[k].kind == "sensitive"
        if k:
            sens_n += 1
            sens_ok += bool(p.get("sensitive")) == is_sens
        rows.append([f"{c:.2f}", p.get("proposition", "")[:60], k or "（对应不上）", {True: "真", False: "假", None: "-"}[truth],
                     "是" if p.get("sensitive") else "否"])
    print_header(f"模型 {name}：从 {len(lines)} 条观察里提出的命题")
    print_table(["置信度", "命题", "对应的隐藏事实", "真假", "它标为敏感"], rows, aligns=[">", "<", "<", "<", "<"])
    print(f"\n能对应上的 {n} 条命题的 Brier 分数：{sq / n:.3f}" if n else "\n没有能对应上的命题")
    print(f"敏感标记和隐藏画像一致：{sens_ok}/{sens_n}")
    print_observations([
        "高置信的命题里有没有假的？它们是被哪几条观察误导的（例如一次点开的广告、一篇顺手打开的文章）？",
        "它有没有写出'评判性'的命题（关于情绪、工作满意度）？GUM 的参与者觉得这类命题'写在纸上看着很怪'。",
        "敏感标记：求职、看心理医生的命题它标了吗？如果它连这些都直接写进了模型，上游的审计（代码 01 演示 4）就是必需的。",
    ])


if __name__ == "__main__":
    main()