tests/test_numeric_router.py

import asyncio

from app.schemas.cross_validation import AnswerKind, TaskType
from app.schemas.numeric import NumericCheckItem
from app.services.numeric_router import run_numeric_checks_with_router


class FakeContext:
    def model_dump(self, mode="json"):
        return {"statement": "Résoudre 3x - 9 = 0."}


class FakeCandidate:
    def __init__(self, task_type, answer_kind):
        self.task_type = task_type
        self.expected_answer_kind = answer_kind


async def fake_translate_equation(**kwargs):
    return FakeCandidate(TaskType.EQUATION, AnswerKind.SOLUTION_SET)


async def fake_translate_derivative(**kwargs):
    return FakeCandidate(TaskType.DERIVATIVE, AnswerKind.EXPRESSION)


async def fake_solution_set_runner(**kwargs):
    return NumericCheckItem(
        check_type="equivalence",
        expression="3",
        expected="x = 3",
        is_valid=True,
        details="student_llm_and_wolfram_are_consistent",
        wolfram_status="ok",
        cas_status="WOLFRAM_ONLY",
        preferred_engine="wolfram",
        confidence_impact="increase",
    )


def test_numeric_router_routes_equation_solution_set(monkeypatch):
    import app.services.numeric_router as router
    monkeypatch.setattr(router, "translate_exercise_to_wolfram_with_llm", fake_translate_equation)
    monkeypatch.setattr(router, "run_solution_set_cross_validation_numeric_check", fake_solution_set_runner)
    report = asyncio.run(run_numeric_checks_with_router(answer_text="S = {3}", pedagogical_context=FakeContext()))
    assert len(report.checks) == 1
    assert report.checks[0].wolfram_status == "ok"
    assert report.checks[0].is_valid is True
    assert report.needs_human_review is False


def test_numeric_router_skips_derivative(monkeypatch):
    import app.services.numeric_router as router
    monkeypatch.setattr(router, "translate_exercise_to_wolfram_with_llm", fake_translate_derivative)
    report = asyncio.run(run_numeric_checks_with_router(answer_text="S = {3}", pedagogical_context=FakeContext()))
    assert len(report.checks) == 1
    assert report.checks[0].wolfram_status == "skipped"
    assert report.checks[0].cas_status == "NO_CAS_VALIDATION"
    assert report.checks[0].is_valid is True
    assert "numeric_router_skipped_unsupported_task=derivative" in report.checks[0].details
    assert report.needs_human_review is False