tests/test_parsing.py

"""Parsing tests — the structured-output contract is the riskiest part with a
small VLM, so it's the part worth testing without a model in the loop.

Run: uv run pytest          (single test: uv run pytest tests/test_parsing.py::test_clean_json)
"""

from tempercheck.prompt import parse_verdict


def test_clean_json():
    v = parse_verdict(
        '{"score": 73, "verdict": "Mildly thorny", '
        '"rationale": "Crossed arms.", "signals": ["frown", "caps lock"]}'
    )
    assert v.score == 73
    assert v.verdict == "Mildly thorny"
    assert v.signals == ["frown", "caps lock"]


def test_json_with_surrounding_prose():
    v = parse_verdict('Sure! Here is the read:\n{"score": 10}\nHope that helps.')
    assert v.score == 10
    assert v.band  # falls back to defaults without raising


def test_score_clamped_and_coerced():
    assert parse_verdict('{"score": 250}').score == 100
    assert parse_verdict('{"score": -5}').score == 0
    assert parse_verdict('{"score": "88 out of 100"}').score == 88


def test_garbage_never_raises():
    v = parse_verdict("the model said nothing useful")
    assert 0 <= v.score <= 100
    assert v.verdict
    assert v.rationale


def test_signals_non_list_is_normalized():
    assert parse_verdict('{"signals": "grumpy"}').signals == ["grumpy"]