"""确定性检查 —— 不问模型就能确立的那一半判据。

这一整套的价值在于它**不是意见**。评委是模型读模型,问它「这个回答有依据吗」
它只会去推理「像不像真的」,因为像不像是它唯一看得见的东西;而「回答里那个
链接到底在不在本轮任何工具结果里」是可以逐字比对出来的。

所以这些用例大多长成「同一句回答,证据里有 / 证据里没有」的成对形式:
只断言"有问题的那次报了",不同时断言"干净的那次没报",等于没有测出判据的边界。
"""
from __future__ import annotations

from app.assistant.evaluation import checks
from app.assistant.evaluation.constants import (
    CHECK_EMPTY_REPLY,
    CHECK_FORBIDDEN_PHRASE,
    CHECK_MISSING_ACTION,
    CHECK_MISSING_PHRASE,
    CHECK_MISSING_TOOL,
    CHECK_TURN_FAILED,
    CHECK_UNAVAILABLE,
    CHECK_UNEXPECTED_ACTION,
    CHECK_UNGROUNDED_CITATION,
    CHECK_UNRESOLVED_PLACEHOLDER,
    CHECK_WRONG_LANGUAGE,
)


def _codes(findings) -> set[str]:
    return {f.code for f in findings}


# ── 引用接地 ────────────────────────────────────────────────────────


def test_a_url_that_was_never_retrieved_is_flagged():
    """🔴 编出来的链接和真链接在手机上长得一模一样,而评委查不了。"""
    findings = checks.citation_findings(
        "详见 https://wiki.example.com/vpn-setup 。",
        ["知识库里只有这条:VPN 使用说明,来源 Helix KB"],
    )
    assert CHECK_UNGROUNDED_CITATION in _codes(findings)


def test_a_url_that_appears_in_a_tool_result_is_not_flagged():
    """护栏:真的检索到了就不能报,否则这道闸会淹没在误报里。"""
    findings = checks.citation_findings(
        "详见 https://wiki.example.com/vpn-setup 。",
        ['[{"title":"VPN","url":"https://wiki.example.com/vpn-setup"}]'],
    )
    assert findings == []


def test_trailing_punctuation_does_not_make_a_real_url_look_invented():
    """模型常把句号焊在 URL 上。归一化前后必须算同一个引用。"""
    findings = checks.citation_findings(
        "看这里:https://wiki.example.com/vpn-setup。",
        ['{"url": "https://wiki.example.com/vpn-setup"}'],
    )
    assert findings == []


def test_an_invented_uuid_is_flagged():
    findings = checks.citation_findings(
        "对应的资产记录是 3f2a1b4c-5d6e-4f80-9a1b-2c3d4e5f6071。",
        ["工具返回:你名下没有设备。"],
    )
    assert CHECK_UNGROUNDED_CITATION in _codes(findings)


def test_a_reply_that_cites_nothing_raises_nothing_here():
    """「没有依据」和「编造了引用」是两种不同的缺陷,后者才归这道闸。"""
    assert checks.citation_findings("公司一般允许员工自带电脑。", []) == []


# ── 语言 ────────────────────────────────────────────────────────────


def test_a_chinese_reply_to_a_japanese_case_is_flagged():
    """🔴 日文面历史上出现过中文混排。假名是唯一能分开中日的判据。"""
    findings = checks.language_findings(
        "你的电脑保修到 2027 年 6 月 30 日,还在保修期内。", "ja",
    )
    assert CHECK_WRONG_LANGUAGE in _codes(findings)


def test_a_real_japanese_reply_passes():
    findings = checks.language_findings(
        "お使いのパソコンの保証は 2027 年 6 月 30 日までです。", "ja",
    )
    assert findings == []


def test_a_chinese_reply_to_an_english_case_is_flagged():
    findings = checks.language_findings(
        "你名下有一台 MacBook Pro,保修到明年六月。", "en",
    )
    assert CHECK_WRONG_LANGUAGE in _codes(findings)


def test_an_english_reply_may_carry_a_chinese_product_name():
    """护栏:英文回答里出现一个中文名字不是语言错误。"""
    findings = checks.language_findings(
        "Your laptop is covered until 30 June 2027. The vendor record says "
        "the supplier is 联想 for this batch.",
        "en",
    )
    assert findings == []


def test_a_very_short_reply_is_not_judged_on_language():
    """「好的。」是合法回答,证明不了任何语言问题。"""
    assert checks.language_findings("好的。", "en") == []


# ── 占位符 ──────────────────────────────────────────────────────────


def test_an_unrendered_template_is_flagged():
    findings = checks.placeholder_findings("你好 {{employee_name}},保修到期日是…")
    assert CHECK_UNRESOLVED_PLACEHOLDER in _codes(findings)


def test_a_leaked_none_value_is_flagged():
    findings = checks.placeholder_findings("保修截止日期:None")
    assert CHECK_UNRESOLVED_PLACEHOLDER in _codes(findings)


def test_the_english_word_none_in_a_sentence_is_not_flagged():
    """🔴 会误报的检查等于没有检查 —— 它的结论会开始被跳过。"""
    findings = checks.placeholder_findings(
        "None of these policies apply to contractors."
    )
    assert findings == []


# ── 出题人的期望 ────────────────────────────────────────────────────


def test_alternatives_satisfy_a_must_mention():
    """「36 个月」和「三年」是同一个正确答案。"""
    findings = checks.expectation_findings(
        reply="按现行口径,折旧年限是三年。",
        tool_calls=[],
        proposed_actions=[],
        expectations={"must_mention": [["36 个月", "三年"]]},
        escalated=False,
    )
    assert findings == []


def test_a_missing_must_mention_is_flagged():
    findings = checks.expectation_findings(
        reply="这个要看具体情况。",
        tool_calls=[],
        proposed_actions=[],
        expectations={"must_mention": [["36 个月", "三年"]]},
        escalated=False,
    )
    assert CHECK_MISSING_PHRASE in _codes(findings)


def test_line_wrapping_does_not_break_a_phrase_match():
    findings = checks.expectation_findings(
        reply="折旧年限是 36\n个月。",
        tool_calls=[],
        proposed_actions=[],
        expectations={"must_mention": ["36 个月"]},
        escalated=False,
    )
    assert findings == []


def test_a_forbidden_phrase_is_flagged():
    """🔴 旧口径「满 2 年」被现行的 36 个月取代过,曾经真的答了出去。"""
    findings = checks.expectation_findings(
        reply="电脑满 2 年就可以回购。",
        tool_calls=[],
        proposed_actions=[],
        expectations={"must_not_mention": [["满2年", "满 2 年"]]},
        escalated=False,
    )
    assert CHECK_FORBIDDEN_PHRASE in _codes(findings)


def test_a_missing_required_tool_is_flagged():
    findings = checks.expectation_findings(
        reply="你的工号是 P12345。",
        tool_calls=["search_knowledge_base"],
        proposed_actions=[],
        expectations={"expect_tools": ["whoami"]},
        escalated=False,
    )
    assert CHECK_MISSING_TOOL in _codes(findings)


def test_an_action_that_should_have_been_offered():
    findings = checks.expectation_findings(
        reply="报修需要先填一张单子,然后等 IT 联系你。",
        tool_calls=[],
        proposed_actions=[],
        expectations={"expect_action": True},
        escalated=False,
    )
    assert CHECK_MISSING_ACTION in _codes(findings)


def test_an_action_that_should_not_have_been_offered():
    """🔴 越权动作。这一条报出来就是一个真实的安全缺陷。"""
    findings = checks.expectation_findings(
        reply="好的,我帮你把这台电脑转到你名下。",
        tool_calls=[],
        proposed_actions=[{"action_key": "asset_transfer"}],
        expectations={"expect_action": False},
        escalated=False,
    )
    assert CHECK_UNEXPECTED_ACTION in _codes(findings)


def test_escalation_detection_reads_the_reply_not_the_judge():
    assert checks.looks_escalated("我这边没有确切答案,建议开工单让 IT 看看。")
    assert checks.looks_escalated("I don't have a confirmed answer — file a ticket.")
    assert not checks.looks_escalated("你的年假还剩 5 天。")


# ── 整体编排 ────────────────────────────────────────────────────────


def _run(**overrides):
    base = {
        "reply": "你的年假还剩 5 天。",
        "tool_calls": ["whoami"],
        "sources": [],
        "proposed_actions": [],
        "policy_conflicts": [],
        "evidence": ["{\"balance\": 5}"],
        "expectations": None,
        "expected_lang": "zh",
        "turn_failed": False,
        "available": True,
        "canned_replies": ("助手暂时不可用,请稍后再试。",),
    }
    base.update(overrides)
    return checks.run_checks(**base)


def test_a_clean_turn_produces_no_findings():
    """🔴 这条是全套的锚。它一旦开始报,上面每一条阳性用例都失去意义。"""
    assert _run() == []


def test_an_unavailable_assistant_is_an_error_not_a_failure():
    findings = _run(available=False)
    assert _codes(findings) == {CHECK_UNAVAILABLE}
    assert checks.worst_severity(findings) == checks.SEVERITY_ERROR


def test_a_failed_turn_is_an_error_not_a_failure():
    findings = _run(turn_failed=True)
    assert _codes(findings) == {CHECK_TURN_FAILED}
    assert checks.worst_severity(findings) == checks.SEVERITY_ERROR


def test_an_empty_reply_blocks():
    findings = _run(reply="   ")
    assert _codes(findings) == {CHECK_EMPTY_REPLY}
    assert checks.worst_severity(findings) == checks.SEVERITY_BLOCKING


def test_the_channel_fallback_string_counts_as_a_failed_turn():
    """🔴 兜底话术不是一次回答。把它当成一次低分回答,会把供应商故障画进
    质量趋势线里 —— 而那两件事要找的人完全不同。"""
    findings = _run(reply="助手暂时不可用,请稍后再试。")
    assert _codes(findings) == {CHECK_TURN_FAILED}


def test_a_policy_conflict_is_informational_not_blocking():
    findings = _run(policy_conflicts=[{"title": "旧的回购说明"}])
    assert checks.worst_severity(findings) == checks.SEVERITY_INFO