PTSRaiAgent/tests/test_rules.py

from datetime import date, datetime, time, timedelta, timezone
from typing import Any

import pytest
from pydantic import ValidationError

from app.core.rules import (
    Action,
    DailyCheckin,
    DoseRecord,
    PatientSnapshot,
    RulesConfig,
    ScalePoint,
    color,
    evaluate,
    load_rules,
    patient_status,
    registered_rule_ids,
)
from app.core.types import AlertLevel

MSK = timezone(timedelta(hours=3), "MSK")
DAY0 = date(2026, 9, 21)  # visit to the psychiatrist, sertraline prescribed


def at(day: int, hour: int = 21) -> datetime:
    return datetime.combine(DAY0 + timedelta(days=day), time(hour), MSK)


def snap(day: int, **kw: Any) -> PatientSnapshot:
    base: dict[str, Any] = {
        "token": "t" * 32,
        "as_of": at(day),
        "age": 28,
        "sex": "m",
        "modules": {"depression"},
        "therapy_day": day,
        "last_patient_message_at": at(day, 20),
    }
    base.update(kw)
    return PatientSnapshot(**base)


def ci(day: int, **kw: Any) -> DailyCheckin:
    return DailyCheckin(day=DAY0 + timedelta(days=day), **kw)


def codes(actions: list[Action], type_: str | None = None) -> set[str]:
    return {a.code for a in actions if type_ is None or a.type == type_}


# ---------------------------------------------------------------------------------------------- config


def test_yaml_loads_and_every_rule_has_evaluator() -> None:
    cfg = load_rules()
    ids = {r.id for r in cfg.rules}
    assert ids == registered_rule_ids()


def test_unknown_rule_id_rejected() -> None:
    with pytest.raises(ValidationError):
        RulesConfig.model_validate({"rules": [{"id": "nope", "actions": [{"type": "alert", "level": "info"}]}]})


def test_alert_without_level_rejected() -> None:
    with pytest.raises(ValidationError):
        RulesConfig.model_validate({"rules": [{"id": "sleep_lt4_3n", "actions": [{"type": "alert"}]}]})


def test_disabled_rule_is_skipped() -> None:
    cfg = load_rules().model_copy(deep=True)
    for r in cfg.rules:
        r.enabled = r.id != "sleep_lt4_3n"
    s = snap(3, checkins=[ci(d, sleep_hours_bucket="<4") for d in (1, 2, 3)])
    assert "sleep_lt4_3n" not in codes(evaluate(s, cfg))


def test_naive_as_of_rejected() -> None:
    with pytest.raises(ValidationError):
        snap(1, as_of=datetime(2026, 9, 22, 21))


def test_quiet_patient_has_no_actions() -> None:
    assert evaluate(snap(3, checkins=[ci(d, mood=6, med_taken=True) for d in (1, 2, 3)])) == []


# ---------------------------------------------------------------------------------------------- crisis rules


def test_phq9_item9_crisis() -> None:
    s = snap(13, scale_results=[ScalePoint(scale="phq9", ts=at(13, 20), total=12, item9=1, scheduled=False)])
    acts = evaluate(s)
    crisis = [a for a in acts if a.code == "phq9_item9"]
    assert [a.type for a in crisis] == ["crisis", "show_safety_plan", "schedule_scale"]
    assert crisis[0].level == AlertLevel.ACUTE
    assert crisis[2].params["scale"] == "cssrs_screen"
    assert "1" in crisis[0].reason_ru


def test_phq9_item9_zero_or_old_is_not_crisis() -> None:
    zero = snap(7, scale_results=[ScalePoint(scale="phq9", ts=at(7, 20), total=13, item9=0)])
    assert "phq9_item9" not in codes(evaluate(zero))
    old = snap(9, scale_results=[ScalePoint(scale="phq9", ts=at(7, 20), total=13, item9=2)])
    assert "phq9_item9" not in codes(evaluate(old))


def test_crisis_never_deduplicated() -> None:
    s = snap(
        13,
        scale_results=[ScalePoint(scale="phq9", ts=at(13, 20), total=12, item9=2)],
        open_alert_codes={"phq9_item9"},
    )
    assert "phq9_item9" in codes(evaluate(s), "crisis")


def test_cssrs_levels() -> None:
    high = evaluate(snap(5, cssrs_risk="high"))
    assert codes(high, "crisis") == {"cssrs_high"}
    moderate = evaluate(snap(5, cssrs_risk="moderate"))
    assert [(a.code, a.level) for a in moderate] == [("cssrs_moderate", AlertLevel.HIGH)]
    assert evaluate(snap(5, cssrs_risk="low")) == []


@pytest.mark.parametrize(
    ("weapon", "cssrs", "item9", "fires"),
    [
        (True, "moderate", None, True),
        (True, None, 1, True),
        (True, "low", 0, False),
        (False, "high", 2, False),
        (None, "moderate", None, False),
    ],
)
def test_weapon_access_with_risk(weapon: bool | None, cssrs: str | None, item9: int | None, fires: bool) -> None:
    scales = [] if item9 is None else [ScalePoint(scale="phq9", ts=at(5, 20), total=10, item9=item9)]
    s = snap(5, weapon_access=weapon, cssrs_risk=cssrs, scale_results=scales)
    assert ("weapon_access_risk" in codes(evaluate(s), "crisis")) is fires


# ---------------------------------------------------------------------------------------------- scale rules


@pytest.mark.parametrize(
    ("scale", "total", "code", "fires"),
    [
        ("phq9", 20, "phq9_severe", True),
        ("phq9", 19, "phq9_severe", False),
        ("pcl5", 50, "pcl5_severe", True),
        ("pcl5", 49, "pcl5_severe", False),
    ],
)
def test_severe_totals(scale: str, total: int, code: str, fires: bool) -> None:
    s = snap(7, scale_results=[ScalePoint(scale=scale, ts=at(7, 10), total=total, item9=0)])
    acts = evaluate(s)
    assert (code in codes(acts)) is fires
    if fires:
        # PCL-5 alone is a planned specialist review, not a priority (clinical review 2026-10-04).
        expected = AlertLevel.MODERATE if scale == "pcl5" else AlertLevel.HIGH
        assert next(a for a in acts if a.code == code).level == expected


@pytest.mark.parametrize(
    ("scale", "prev", "last", "code", "fires"),
    [
        ("phq9", 10, 15, "phq9_rise", True),
        ("phq9", 10, 14, "phq9_rise", False),
        ("pcl5", 30, 40, "pcl5_rise", True),
        ("pcl5", 30, 39, "pcl5_rise", False),
        ("pcl5", 40, 30, "pcl5_rise", False),
    ],
)
def test_rise_between_last_two(scale: str, prev: int, last: int, code: str, fires: bool) -> None:
    s = snap(
        14,
        scale_results=[
            ScalePoint(scale=scale, ts=at(14, 10), total=last, item9=0),  # order in input does not matter
            ScalePoint(scale=scale, ts=at(0, 10), total=prev, item9=0),
        ],
    )
    assert (code in codes(evaluate(s))) is fires


def test_future_scale_results_ignored() -> None:
    s = snap(7, scale_results=[ScalePoint(scale="phq9", ts=at(8), total=25, item9=3)])
    assert evaluate(s) == []


@pytest.mark.parametrize(("sex", "total", "fires"), [("m", 4, True), ("m", 3, False), ("f", 3, True), ("f", 2, False)])
def test_auditc_by_sex(sex: str, total: int, fires: bool) -> None:
    s = snap(1, sex=sex, scale_results=[ScalePoint(scale="auditc", ts=at(0), total=total)])
    acts = evaluate(s)
    assert ("auditc_positive" in codes(acts)) is fires
    if fires:
        assert acts[0].level == AlertLevel.MODERATE


# ---------------------------------------------------------------------------------------------- adherence


def test_quit_intent_voiced() -> None:
    acts = evaluate(snap(5, checkins=[ci(5, med_taken=False, quit_intent=True)]))
    assert [(a.type, a.level) for a in acts if a.code == "quit_intent"] == [
        ("alert", AlertLevel.HIGH),
        ("med_info", None),
    ]
    assert next(a for a in acts if a.type == "med_info").params["topic"] == "delayed_effect"


def test_quit_intent_followed_by_skipped_dose() -> None:
    s = snap(7, checkins=[ci(5, med_taken=True, quit_intent=True), ci(6, med_taken=True), ci(7, med_taken=False)])
    assert "quit_intent" in codes(evaluate(s), "alert")
    no_intent = snap(7, checkins=[ci(6, med_taken=True), ci(7, med_taken=False)])
    assert "quit_intent" not in codes(evaluate(no_intent))


def test_missed_doses_3_of_7() -> None:
    doses = [
        DoseRecord(planned_at=at(d, 9), status=st)
        for d, st in [(1, "missed"), (2, "taken"), (3, "skipped"), (4, "taken"), (5, "taken"), (6, "taken")]
    ]
    s = snap(7, doses=doses, checkins=[ci(7, med_taken=False)])
    acts = [a for a in evaluate(s) if a.code == "missed_doses_7d"]
    assert [a.type for a in acts] == ["alert", "reach_out"]
    assert acts[0].level == AlertLevel.MODERATE and acts[0].params["count"] == 3
    # Day 1 slides out of the 7-day window on day 8 → only 2 missed days left.
    assert "missed_doses_7d" not in codes(evaluate(snap(8, doses=doses, checkins=[ci(7, med_taken=False)])))


def test_side_effects_three_consecutive_days() -> None:
    two = snap(3, checkins=[ci(1), ci(2, side_effects=["nausea"]), ci(3, side_effects=["nausea"])])
    assert evaluate(two) == []
    gap = snap(
        4,
        checkins=[
            ci(1, side_effects=["nausea"]),
            ci(3, side_effects=["nausea"]),
            ci(4, side_effects=["nausea"]),
        ],
    )
    assert evaluate(gap) == []
    three = snap(4, checkins=[ci(d, side_effects=["nausea"]) for d in (2, 3, 4)])
    acts = evaluate(three)
    assert [(a.code, a.level) for a in acts] == [("side_effect_3d", AlertLevel.MODERATE)]
    assert "nausea" in acts[0].reason_ru


def test_streak_seen_in_morning_before_todays_checkin() -> None:
    s = snap(5, as_of=at(5, 8), checkins=[ci(d, side_effects=["nausea"]) for d in (2, 3, 4)])
    assert "side_effect_3d" in codes(evaluate(s))
    stale = snap(6, as_of=at(6, 8), checkins=[ci(d, side_effects=["nausea"]) for d in (2, 3, 4)])
    assert evaluate(stale) == []


# ---------------------------------------------------------------------------------- check-in patterns


def test_sleep_lt4_three_nights() -> None:
    s = snap(10, checkins=[ci(d, sleep_hours_bucket="<4") for d in (8, 9, 10)])
    assert [(a.code, a.level) for a in evaluate(s)] == [("sleep_lt4_3n", AlertLevel.MODERATE)]
    s2 = snap(
        10,
        checkins=[
            ci(8, sleep_hours_bucket="<4"),
            ci(9, sleep_hours_bucket="4-6"),
            ci(10, sleep_hours_bucket="<4"),
        ],
    )
    assert evaluate(s2) == []


def test_alcohol_three_days() -> None:
    s = snap(10, checkins=[ci(8, alcohol="1-2"), ci(9, alcohol="more"), ci(10, alcohol="more")])
    acts = evaluate(s)
    assert [a.type for a in acts] == ["alert", "med_info", "suggest_visit"]
    assert acts[0].level == AlertLevel.MODERATE
    assert acts[1].params["topic"] == "alcohol_interaction"
    s2 = snap(10, checkins=[ci(8, alcohol="1-2"), ci(9, alcohol="none"), ci(10, alcohol="more")])
    assert evaluate(s2) == []


def test_low_mood_schedules_unscheduled_phq9_with_cooldown() -> None:
    checkins = [ci(11, mood=3), ci(12, mood=3), ci(13, mood=2)]
    acts = evaluate(snap(13, checkins=checkins))
    assert [(a.type, a.level, a.params) for a in acts] == [
        ("schedule_scale", None, {"scale": "phq9", "scheduled": False})
    ]
    done = ScalePoint(scale="phq9", ts=at(13, 20), total=12, item9=0, scheduled=False)
    assert evaluate(snap(13, checkins=checkins, scale_results=[done])) == []
    assert evaluate(snap(13, checkins=[ci(11, mood=3), ci(12, mood=4), ci(13, mood=2)])) == []


def test_silence_over_3_days() -> None:
    acts = evaluate(snap(10, last_patient_message_at=at(6, 20)))
    assert [(a.type, a.level) for a in acts] == [("alert", AlertLevel.INFO), ("reach_out", None)]
    assert acts[0].params["silent_days"] == 4
    assert evaluate(snap(10, last_patient_message_at=at(7, 21))) == []  # exactly 3 days


@pytest.mark.parametrize(("age", "day", "fires"), [(23, 8, True), (24, 28, True), (25, 8, False), (23, 29, False)])
def test_young_adult_enhanced_monitoring(age: int, day: int, fires: bool) -> None:
    acts = evaluate(snap(day, age=age))
    assert (acts == [a for a in acts if a.type == "enhanced_monitoring"] and bool(acts)) is fires
    if fires:
        assert acts[0].level == AlertLevel.INFO


# ---------------------------------------------------------------------------------------------- dedup


def test_open_alert_suppresses_whole_rule() -> None:
    checkins = [ci(8, alcohol="1-2"), ci(9, alcohol="more"), ci(10, alcohol="more")]
    assert evaluate(snap(10, checkins=checkins, open_alert_codes={"alcohol_3d"})) == []
    # Other rules are not affected by an unrelated open code.
    assert codes(evaluate(snap(10, checkins=checkins, open_alert_codes={"sleep_lt4_3n"}))) == {"alcohol_3d"}


def test_evaluate_is_deterministic() -> None:
    s = snap(10, checkins=[ci(d, alcohol="more", sleep_hours_bucket="<4") for d in (8, 9, 10)])
    assert evaluate(s) == evaluate(s)


# ---------------------------------------------------------------------------------------------- colours


@pytest.mark.parametrize(
    ("levels", "expected"),
    [
        ([], "green"),
        ([AlertLevel.INFO], "grey"),
        ([AlertLevel.INFO, AlertLevel.MODERATE], "yellow"),
        ([AlertLevel.HIGH], "yellow"),
        ([AlertLevel.MODERATE, AlertLevel.ACUTE], "red"),
    ],
)
def test_color(levels: list[AlertLevel], expected: str) -> None:
    assert color(levels) == expected


def test_patient_status() -> None:
    st = patient_status([AlertLevel.MODERATE, None, AlertLevel.HIGH])
    assert st.color == "yellow" and st.top_level == AlertLevel.HIGH
    assert patient_status([]).top_level is None


# ---------------------------------------------------------------------------------------------- demo persona
# docs/product/01_demo_persona.md, scenario B: Alexey, 28, sertraline 50 mg from day 0.
# (day, med_taken, mood, sleep, alcohol, side effects, quit_intent)
PERSONA_DAYS: list[tuple[int, bool, int, str, str, list[str], bool]] = [
    (1, True, 4, "6-8", "none", [], False),
    (2, True, 4, "4-6", "none", ["nausea"], False),
    (3, True, 4, "4-6", "none", ["nausea"], False),
    (4, True, 3, "4-6", "none", ["nausea"], False),
    (5, False, 3, "4-6", "none", [], True),  # «от таблеток только хуже, брошу»
    (6, True, 4, "6-8", "none", [], False),  # doctor called, patient continues
    (7, True, 4, "6-8", "none", [], False),
    (8, True, 4, "<4", "1-2", [], False),
    (9, True, 4, "<4", "more", [], False),
    (10, True, 4, "<4", "more", [], False),  # «не могу уснуть без пива»
    (11, True, 3, "4-6", "1-2", [], False),
    (12, True, 3, "4-6", "none", [], False),
    (13, True, 2, "<4", "none", [], False),  # «всё бессмысленно, устал от всего»
]
PERSONA_SCALES = [
    ScalePoint(scale="phq9", ts=at(-3, 12), total=14, item9=0, scheduled=True),  # screening
    ScalePoint(scale="phq9", ts=at(7, 19), total=13, item9=0, scheduled=True),  # weekly PHQ-9
    ScalePoint(scale="phq9", ts=at(13, 20), total=18, item9=2, scheduled=False),  # unscheduled (mood ≤ 3 × 3)
]
EXPECTED = ["G", "G", "G", "Y", "Y", "G", "G", "G", "G", "Y", "Y", "Y", "R"]
COLOR_LETTER = {"green": "G", "yellow": "Y", "red": "R", "grey": "g"}
DOCTOR_ACK_DAY = 6  # the doctor called the patient and acked the open alerts


def _persona_snapshot(day: int, open_codes: set[str], include_scales_until: datetime) -> PatientSnapshot:
    rows = [r for r in PERSONA_DAYS if r[0] <= day]
    return snap(
        day,
        checkins=[
            ci(
                d,
                med_taken=med,
                mood=mood,
                sleep_hours_bucket=sleep,
                alcohol=alc,
                side_effects=se,
                quit_intent=qi,
            )
            for d, med, mood, sleep, alc, se, qi in rows
        ],
        doses=[DoseRecord(planned_at=at(d, 9), status="taken" if med else "skipped") for d, med, *_ in rows],
        scale_results=[p for p in PERSONA_SCALES if p.ts <= include_scales_until],
        open_alert_codes=set(open_codes),
    )


def test_demo_persona_trajectory() -> None:
    open_alerts: dict[str, AlertLevel] = {}
    colours: list[str] = []
    new_alerts: dict[int, set[str]] = {}
    for day in range(1, 14):
        if day == DOCTOR_ACK_DAY:
            open_alerts.clear()  # quit_intent + side_effect_3d acked after the call
        if day == 13:
            # Before the unscheduled PHQ-9: mood ≤ 3 for 3 days triggers it.
            pre = evaluate(_persona_snapshot(day, set(open_alerts), include_scales_until=at(13, 19)))
            assert codes(pre, "schedule_scale") == {"mood_low_3d"}
        acts = evaluate(_persona_snapshot(day, set(open_alerts), include_scales_until=at(day, 23)))
        fresh = {a.code: a.level for a in acts if a.creates_alert and a.level is not None}
        new_alerts[day] = set(fresh)
        open_alerts.update(fresh)
        colours.append(COLOR_LETTER[color(open_alerts.values())])

    assert colours == EXPECTED
    assert new_alerts[4] == {"side_effect_3d"}
    assert new_alerts[5] == {"quit_intent"}
    assert new_alerts[10] == {"sleep_lt4_3n", "alcohol_3d"}
    assert new_alerts[11] == set() and new_alerts[12] == set()  # still open, not re-emitted
    assert new_alerts[13] == {"phq9_item9", "phq9_rise"}
    assert {d for d, c in new_alerts.items() if c} == {4, 5, 10, 13}