import json

import numpy as np

import signspeak.asl.asl_detector as asl_detector_module
from signspeak.asl.asl_detector import ASLDetector


def test_load_labels_prefers_prediction_index_map(tmp_path):
    model_dir = tmp_path / "asl"
    model_dir.mkdir()
    (model_dir / "sign_to_prediction_index_map.json").write_text(
        json.dumps({"love": 2, "hello": 0, "thanks": 1}),
        encoding="utf-8",
    )

    detector = ASLDetector(model_dir=model_dir)

    assert detector.labels == ["hello", "thanks", "love"]


class FakeSignatureInterpreter:
    def get_signature_list(self):
        return {"serving_default": {"inputs": ["inputs"], "outputs": ["outputs"]}}

    def get_signature_runner(self, signature_name):
        assert signature_name == "serving_default"

        def predict(**kwargs):
            assert kwargs["inputs"].shape == (543, 3)
            assert np.isnan(kwargs["inputs"][0][0])
            return {"outputs": np.asarray([[0.1, 0.8, 0.1]], dtype=np.float32)}

        return predict


def test_predict_uses_tflite_signature_runner(tmp_path):
    detector = ASLDetector(model_dir=tmp_path)
    keypoints = np.zeros((543, 3), dtype=np.float32)
    keypoints[0][0] = np.nan

    output = detector._predict(FakeSignatureInterpreter(), keypoints)

    assert output.shape == (1, 3)
    assert float(output[0][1]) == np.float32(0.8)


class FakeLandmarkResult:
    keypoints = np.zeros((30, 543, 3), dtype=np.float32)
    status = "ok"
    detector = "fake"
    error = None


class FakeLandmarksDetector:
    def __init__(self, missing_value=0.0):
        pass

    def detect_sequence(self, frames):
        return FakeLandmarkResult()

    def close(self):
        pass


def test_low_confidence_prediction_is_not_accepted(monkeypatch, tmp_path):
    model_dir = tmp_path / "asl"
    model_dir.mkdir()
    (model_dir / "model.tflite").write_bytes(b"demo")
    (model_dir / "sign_to_prediction_index_map.json").write_text(
        json.dumps({"where": 0, "hello": 1}),
        encoding="utf-8",
    )
    monkeypatch.setattr(asl_detector_module, "LandmarksDetector", FakeLandmarksDetector)
    monkeypatch.setenv("ASL_CONFIDENCE_THRESHOLD", "0.70")
    monkeypatch.setattr(ASLDetector, "_load_interpreter", lambda self: object())
    monkeypatch.setattr(
        ASLDetector,
        "_predict",
        lambda self, interpreter, keypoints: np.asarray([[0.667, 0.333]], dtype=np.float32),
    )

    result = ASLDetector(model_dir=model_dir).predict_from_frames([np.zeros((2, 2, 3), dtype=np.uint8)] * 30)

    assert result["status"] == "low_confidence"
    assert result["top_prediction"] == "where"
    assert result["gloss_sequence"] == []


def test_sequence_prediction_collapses_accepted_windows(monkeypatch, tmp_path):
    model_dir = tmp_path / "asl"
    model_dir.mkdir()
    (model_dir / "model.tflite").write_bytes(b"demo")
    (model_dir / "sign_to_prediction_index_map.json").write_text(
        json.dumps({"hello": 0, "water": 1}),
        encoding="utf-8",
    )

    class SequenceLandmarkResult:
        keypoints = np.zeros((60, 543, 3), dtype=np.float32)
        status = "ok"
        detector = "fake"
        error = None

    class SequenceLandmarksDetector:
        def __init__(self, missing_value=0.0):
            pass

        def detect_sequence(self, frames):
            return SequenceLandmarkResult()

        def close(self):
            pass

    calls = iter(
        [
            np.asarray([[0.9, 0.1]], dtype=np.float32),
            np.asarray([[0.86, 0.14]], dtype=np.float32),
            np.asarray([[0.1, 0.9]], dtype=np.float32),
        ]
    )
    monkeypatch.setattr(asl_detector_module, "LandmarksDetector", SequenceLandmarksDetector)
    monkeypatch.setenv("ASL_CONFIDENCE_THRESHOLD", "0.70")
    monkeypatch.setenv("ASL_SEQUENCE_WINDOW", "30")
    monkeypatch.setenv("ASL_SEQUENCE_STRIDE", "15")
    monkeypatch.setattr(ASLDetector, "_load_interpreter", lambda self: object())
    monkeypatch.setattr(ASLDetector, "_predict", lambda self, interpreter, keypoints: next(calls))

    result = ASLDetector(model_dir=model_dir).predict_sequence_from_frames(
        [np.zeros((2, 2, 3), dtype=np.uint8)] * 60
    )

    assert result["status"] == "ok"
    assert result["gloss_sequence"] == ["hello", "water"]
    assert result["windows_used"] == 3
    assert result["recognition_mode"] == "sliding_window_sequence"