"""助手现在够得到安全答案库了 —— 两个入口一起。

这条洞的形状很具体:答案库的 71 条全部是客户安全问卷的口吻,而员工问的是
「我能用个人电脑写公司代码吗」。员工端的安全 FAQ 页为此长出了混合检索,而**助手
根本答不了这个问题** —— 它会给一件公司早就写下来的事开一张工单。

工具装在共享内核(`copilot.TOOLS`)上而不是飞书那一侧,所以网页 IT 求助和飞书
bot 同时拿到。这一批钉住的是:

* 两个入口都真的看得到它(装在共享内核上,不是某一个渠道的 extra_tools);
* 它**严格只读** —— 一个模型一轮里可能调它三次,而一个「读」工具去写库正是本仓
  savepoint 闸存在的那个形状;
* 它只返回员工可见的条目;
* 两边的提示词都告诉了模型它存在,而且「没有依据就明说」这条对安全问题最严格。
"""
from __future__ import annotations

import json
import uuid

import pytest

from app.domains.support.copilot import SECURITY_TOOL_NAME

pytestmark = pytest.mark.asyncio


async def _entry(db, *, visible: bool = True, zh: str | None = None):
    from app.domains.auditor.models import AuditorAnswerLibraryEntry

    row = AuditorAnswerLibraryEntry(
        id=uuid.uuid4(),
        category="general",
        question_zh=zh or "贵公司是否允许员工使用个人设备处理公司数据?",
        question_en="Does your company permit BYOD for company data?",
        answer_zh="公司允许在满足 MDM 纳管的前提下使用个人设备处理公司数据。",
        answer_en="BYOD is permitted when the device is MDM-enrolled.",
        status="active",
        employee_visible=visible,
    )
    db.add(row)
    await db.flush()
    return row


async def _person(db):
    from app.identity.person_models import IdentityPerson

    person = IdentityPerson(
        id=uuid.uuid4(),
        source_system="successfactors",
        hris_person_id=f"P{uuid.uuid4().hex[:8]}",
        display_name="提问的员工",
        work_email=f"u{uuid.uuid4().hex[:8]}@pingcap.com",
        employment_status="active",
    )
    db.add(person)
    await db.flush()
    return person


# ── 装在哪 ──────────────────────────────────────────────────────────


def test_the_tool_lives_on_the_shared_kernel():
    """🔴 两个入口一起吃到,是这条洞的修法本身。

    装在飞书那一侧的 ``extra_tools`` 上,网页 IT 求助就仍然答不了安全问题 ——
    而那是同一个问题的另一半。
    """
    from app.domains.support.copilot import TOOLS

    names = [tool["name"] for tool in TOOLS]
    assert SECURITY_TOOL_NAME in names


def test_both_prompts_tell_the_model_it_exists():
    """一个模型不知道存在的工具等于不存在。"""
    from app.assistant.prompt import feishu_system_prompt
    from app.domains.support.copilot import SYSTEM_PROMPT

    feishu = feishu_system_prompt("zh")
    assert SECURITY_TOOL_NAME in feishu
    # 网页那一份没有把工具名插进去(它的措辞是「the security-policy tool」),
    # 所以这里断言的是意图而不是字面。
    assert "security-policy tool" in SYSTEM_PROMPT
    assert "on the company's behalf" in SYSTEM_PROMPT


def test_the_feishu_prompt_routes_permission_questions_to_it():
    """「能不能…」是权限问题,不是操作问题 —— 路由顺序里要认得出来。"""
    from app.assistant.prompt import feishu_system_prompt

    text = feishu_system_prompt("zh")
    assert "POLICY question" in text
    # 而且要说清它和知识库不是同一件事,否则模型会以为查过 KB 就够了。
    assert "a how-to article is not a policy statement" in text


# ── 行为 ────────────────────────────────────────────────────────────


async def test_it_answers_from_the_library(db_session):
    from app.domains.support.copilot import _dispatch_tool

    await _entry(db_session)
    person = await _person(db_session)
    await db_session.commit()

    raw = await _dispatch_tool(
        SECURITY_TOOL_NAME,
        {"question": "个人设备处理公司数据可以吗"},
        db_session,
        person,
    )
    payload = json.loads(raw)
    assert payload["covered"] is True
    assert payload["entries"], "没有召回到任何条目"
    assert "MDM" in payload["entries"][0]["answer"]


async def test_it_never_returns_an_entry_employees_cannot_see(db_session):
    """🔴 这个工具是员工在用,答案库里有不对员工公开的条目。"""
    from app.domains.support.copilot import _dispatch_tool

    await _entry(db_session, visible=False, zh="贵公司的内部渗透测试频率?")
    person = await _person(db_session)
    await db_session.commit()

    raw = await _dispatch_tool(
        SECURITY_TOOL_NAME,
        {"question": "内部渗透测试频率"},
        db_session,
        person,
    )
    payload = json.loads(raw)
    assert payload["entries"] == []
    assert payload["covered"] is False


async def test_it_writes_nothing(db_session):
    """🔴 严格只读,而且这是一个决定而不是疏漏。

    ``ask(record_gaps=True)`` 会往安全团队的缺口台账写一行 —— 但一个模型在一轮里
    可能换着措辞调这个工具三次,那会把同一个缺口记三遍;而一个「读」工具去写库,
    正是本仓 savepoint 闸存在的那个形状:这里一次失败的 insert 会在任何东西解释它
    之前若干条语句就把整轮的事务毒死。
    """
    from sqlalchemy import func, select

    from app.domains.auditor.models import AuditorSecurityFaqGap
    from app.domains.support.copilot import _dispatch_tool

    await _entry(db_session)
    person = await _person(db_session)
    await db_session.commit()
    before = (
        await db_session.execute(
            select(func.count()).select_from(AuditorSecurityFaqGap)
        )
    ).scalar()

    # 这句话必须**一条都召回不到**,那才是 record_gaps=True 会写行的那条路。
    # 第一版用的是「公司给员工报销家庭宽带吗」—— 它和库里那条共享「公司」「员工」
    # 两个词,于是清过了词项线、走进了「召回到了」的分支,而那条分支只在模型说
    # 「都答不上」时才写行(本地没配模型 → 不写)。也就是说那一版无论 record_gaps
    # 是真是假都不写行,断言是空转的。
    probe = "熊猫会不会飞"
    from app.domains.auditor import security_faq_service as security

    precheck = await security.ask(
        db_session, query=probe, lang="zh", person_id=None,
        limit=1, compose=False, record_gaps=False,
    )
    assert not precheck.get("items"), (
        "前置条件不成立:这句探针被召回到了,走不到会写缺口的那条分支"
    )

    await _dispatch_tool(
        SECURITY_TOOL_NAME, {"question": probe}, db_session, person,
    )
    await db_session.commit()
    after = (
        await db_session.execute(
            select(func.count()).select_from(AuditorSecurityFaqGap)
        )
    ).scalar()
    assert after == before, "只读工具往缺口台账写了行"


async def test_an_uncovered_question_says_so_rather_than_returning_nothing(
    db_session,
):
    """`covered: false` 是一个可用的答案 —— 模型要能据此说「库里没有」。"""
    from app.domains.support.copilot import _dispatch_tool

    person = await _person(db_session)
    await db_session.commit()

    raw = await _dispatch_tool(
        SECURITY_TOOL_NAME, {"question": "公司允许养宠物进办公室吗"}, db_session, person,
    )
    payload = json.loads(raw)
    assert payload["covered"] is False
    assert "does not cover it" in payload["note"]


async def test_an_empty_question_is_refused_without_a_query(db_session):
    from app.domains.support.copilot import _dispatch_tool

    person = await _person(db_session)
    await db_session.commit()
    payload = json.loads(
        await _dispatch_tool(SECURITY_TOOL_NAME, {"question": "  "}, db_session, person)
    )
    assert "error" in payload


async def test_it_hands_the_model_answers_and_not_retrieval_internals(db_session):
    """🔴 把置信度分层、通道来源递给模型,它会把检索内部状态当政策念给员工。"""
    from app.domains.support.copilot import _dispatch_tool

    await _entry(db_session)
    person = await _person(db_session)
    await db_session.commit()

    payload = json.loads(
        await _dispatch_tool(
            SECURITY_TOOL_NAME,
            {"question": "个人设备处理公司数据可以吗"},
            db_session,
            person,
        )
    )
    assert set(payload) == {"covered", "lead_answer", "entries", "note"}
    assert set(payload["entries"][0]) == {"id", "question", "answer"}