Gerador de vídeos narrativos (Godot + FFmpeg), painel Next.js, serviço de voz, Clipador de lives e Publicador para YouTube, TikTok e Instagram. Inclui a leitura das visualizações dos posts para o Dashboard e o botão "Publicar no <rede>" nos cartões das publicações. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
197 lines
8.6 KiB
Python
197 lines
8.6 KiB
Python
import random
|
|
|
|
from app.scoring import ChatMessage, Params, Word, find_moments, message_weight, overlaps, snap_to_words
|
|
from app.subtitles import build_ass, group_words
|
|
|
|
|
|
def calm_chat(duration: int, rate: float = 3.0, seed: int = 1) -> list[ChatMessage]:
|
|
rng = random.Random(seed)
|
|
out = []
|
|
for t in range(duration):
|
|
for _ in range(int(rate) + (1 if rng.random() < rate % 1 else 0)):
|
|
out.append(ChatMessage(t + rng.random(), rng.choice(["oi", "salve", "boa noite", "eae", "kappa"])))
|
|
return out
|
|
|
|
|
|
def burst(at: int, seconds: int, per_s: int, text: str) -> list[ChatMessage]:
|
|
return [ChatMessage(at + s + i / per_s, text) for s in range(seconds) for i in range(per_s)]
|
|
|
|
|
|
def test_message_weight_counts_reactions():
|
|
assert message_weight("oi")[0] == 1.0
|
|
weight, tags = message_weight("KKKKKKK clipa isso")
|
|
assert "risada" in tags and "clip" in tags and weight == 5.0
|
|
|
|
|
|
def test_finds_peak_after_the_event_and_cuts_before_it():
|
|
chat = calm_chat(600) + burst(300, 8, 25, "KKKKKKKKK")
|
|
moments = find_moments(600, chat, [], [])
|
|
assert len(moments) == 1
|
|
m = moments[0]
|
|
assert 300 <= m.peak_s <= 310
|
|
assert m.start_s < 300 - 5 # o corte começa ANTES da reação (é ali que aconteceu)
|
|
assert m.end_s >= m.peak_s
|
|
assert Params().min_s <= m.end_s - m.start_s <= Params().max_s
|
|
assert m.source == "chat" and m.reasons["tags"].get("risada")
|
|
|
|
|
|
def test_big_chat_needs_relative_peak():
|
|
# Chat enorme o tempo todo (canal grande) não vira corte sem um pico relativo.
|
|
assert find_moments(600, calm_chat(600, rate=40), [], []) == []
|
|
|
|
|
|
def test_two_peaks_far_apart_both_found_and_close_ones_merged():
|
|
chat = calm_chat(900) + burst(200, 6, 25, "caralho") + burst(215, 4, 25, "kkkkk") + burst(600, 6, 25, "POGGERS")
|
|
peaks = [m.peak_s for m in find_moments(900, chat, [], [])]
|
|
assert len(peaks) == 2
|
|
assert 195 <= peaks[0] <= 225 and 595 <= peaks[1] <= 610
|
|
|
|
|
|
def test_live_mode_waits_for_settle():
|
|
chat = calm_chat(316) + burst(300, 8, 25, "KKKK")
|
|
assert find_moments(316, chat, [], [], live=True) == [] # pico perto demais do fim da gravação
|
|
chat = calm_chat(400) + burst(300, 8, 25, "KKKK")
|
|
assert len(find_moments(400, chat, [], [], live=True)) == 1
|
|
|
|
|
|
def test_live_mode_does_not_cut_a_burst_that_is_still_growing():
|
|
# Explosão começa em 300 e só atinge o máximo em 312; com 326 s gravados o máximo ainda não assentou,
|
|
# e o começo da explosão (já "assentado") não pode virar corte sozinho.
|
|
chat = calm_chat(326) + burst(300, 6, 10, "kkk") + burst(306, 8, 40, "KKKKKKK")
|
|
assert find_moments(326, chat, [], [], live=True) == []
|
|
|
|
|
|
def test_audio_mode_without_chat():
|
|
loud = [-40.0] * 600
|
|
for t in range(400, 406):
|
|
loud[t] = -8.0
|
|
moments = find_moments(600, [], loud, [])
|
|
assert len(moments) == 1 and moments[0].source == "audio" and 395 <= moments[0].peak_s <= 410
|
|
|
|
|
|
def test_snap_to_words_avoids_cutting_mid_sentence():
|
|
words = [Word(10.0, 10.4, "olha"), Word(10.5, 10.9, "isso"), Word(12.0, 12.3, "não"), Word(12.35, 12.8, "acredito"),
|
|
Word(20.0, 20.5, "mano")]
|
|
start, end = snap_to_words(12.2, 12.5, words)
|
|
assert start <= 12.0 and end >= 12.8
|
|
|
|
|
|
def test_overlaps():
|
|
assert overlaps(0, 30, 20, 50)
|
|
assert not overlaps(0, 30, 27, 60) # tolerância de 5 s
|
|
|
|
|
|
def test_ass_groups_words_and_escapes_tags():
|
|
words = [Word(1.0, 1.3, "olha"), Word(1.35, 1.6, "{isso}"), Word(1.7, 2.0, "aqui"), Word(5.0, 5.4, "gente")]
|
|
groups = group_words(words)
|
|
assert [len(g) for g in groups] == [3, 1]
|
|
ass = build_ass(words, 0.5, 6.0, "twitch.tv/alanzoka")
|
|
assert "Credit" in ass and "twitch.tv/alanzoka" in ass
|
|
assert "{isso}" not in ass and "(ISSO)" in ass # chave do texto falado não vira tag ASS
|
|
assert "0:00:00.50" in ass # "olha" em 1.0 s da sessão = 0.5 s do corte
|
|
|
|
|
|
def test_params_default_follows_cut_settings_defaults():
|
|
# Pedido do dono (16/11/2026): cortes de 60 a 90 s por padrão — a heurística segue a configuração do corte.
|
|
from app.layout import DEFAULTS
|
|
assert (Params().min_s, Params().max_s) == (DEFAULTS["duration"]["minS"], DEFAULTS["duration"]["maxS"]) == (60, 90)
|
|
|
|
|
|
def test_regression_calm_chat_has_no_false_positives():
|
|
# Medido na live real (16/11/2026): ~0,6 msg/s, 3 mensagens com um emote "clip5k1hmm" viravam corte,
|
|
# e o começo da gravação também.
|
|
chat = calm_chat(180, rate=0.6)
|
|
chat += [ChatMessage(100.2, "l i b e r t e s é!"), ChatMessage(100.5, "clip5k1hmm"), ChatMessage(101.0, "ff")]
|
|
chat += [ChatMessage(3 + i * 0.3, "130") for i in range(5)]
|
|
assert message_weight("clip5k1hmm")[0] == 1.0
|
|
assert message_weight("CLIP IT")[1] == ["clip"]
|
|
assert find_moments(180, chat, [], []) == []
|
|
|
|
|
|
def test_ass_keeps_tags_lowercase_when_uppercasing_text():
|
|
ass = build_ass([Word(1.0, 1.3, "olha"), Word(1.35, 1.6, "isso")], 0.0, 5.0, "")
|
|
assert r"{\c&H00E5FF&}OLHA{\c&HFFFFFF&} ISSO" in ass and r"\C&" not in ass
|
|
|
|
|
|
def test_regression_sparse_chat_still_finds_reactions():
|
|
# Medido em 17/11/2026 (skipnho): 0,15 msg/s; reação de 6 mensagens em 8 s nunca passava do mínimo fixo de 12.
|
|
chat = calm_chat(900, rate=0.15, seed=3)
|
|
chat += [ChatMessage(500 + i, t) for i, t in enumerate(["KKKKKK", "kkkkkkk", "caralho", "KKKKK", "que isso", "kkkk"])]
|
|
moments = [m for m in find_moments(900, chat, [], []) if m.source == "chat"]
|
|
assert len(moments) == 1 and 495 <= moments[0].peak_s <= 510
|
|
|
|
|
|
def test_sparse_chat_also_uses_audio():
|
|
chat = calm_chat(600, rate=0.15, seed=4)
|
|
loud = [-40.0] * 600
|
|
for t in range(300, 306):
|
|
loud[t] = -6.0
|
|
moments = find_moments(600, chat, loud, [])
|
|
assert any(m.source == "audio" and 295 <= m.peak_s <= 312 for m in moments)
|
|
|
|
|
|
# ---------------------------------------------------------------- rosto do streamer (reaction.py)
|
|
from app.reaction import FaceSecond # noqa: E402
|
|
|
|
|
|
def calm_face(n: int, seed: int = 11) -> list[FaceSecond]:
|
|
rng = random.Random(seed)
|
|
return [FaceSecond(t, 1.0, 0.1, 0.5, 0.15, smile=rng.uniform(0.02, 0.22), jaw=rng.uniform(0.0, 0.01),
|
|
surprise=0.002, frown=rng.uniform(0.01, 0.07), brow=rng.uniform(0.01, 0.05), blink=0.1,
|
|
motion=rng.uniform(0.02, 0.15)) for t in range(n)]
|
|
|
|
|
|
def laugh_at(face: list[FaceSecond], start: int, seconds: int = 6) -> list[FaceSecond]:
|
|
for t in range(start, start + seconds):
|
|
face[t].smile = 0.9
|
|
return face
|
|
|
|
|
|
def noisy_loudness(n: int, seed: int = 2) -> list[float]:
|
|
rng = random.Random(seed)
|
|
return [-30.0 + rng.uniform(-1.5, 1.5) for _ in range(n)]
|
|
|
|
|
|
def test_face_none_or_empty_keeps_old_behavior():
|
|
chat = calm_chat(600) + burst(300, 8, 25, "KKKKKKKKK")
|
|
loud = noisy_loudness(600)
|
|
base = find_moments(600, chat, loud, [])
|
|
for face in (None, [], [FaceSecond(t) for t in range(600)]): # sem rosto em nenhum segundo
|
|
again = find_moments(600, chat, loud, [], face=face)
|
|
assert [(m.peak_s, m.start_s, m.end_s, m.score, m.source, m.reasons) for m in again] == \
|
|
[(m.peak_s, m.start_s, m.end_s, m.score, m.source, m.reasons) for m in base]
|
|
|
|
|
|
def test_face_reinforces_chat_moment_with_limited_bonus():
|
|
chat = calm_chat(600) + burst(300, 8, 25, "KKKKKKKKK")
|
|
base = find_moments(600, chat, [], [])[0]
|
|
m = find_moments(600, chat, [], [], face=laugh_at(calm_face(600), 296))[0]
|
|
assert (m.peak_s, m.start_s, m.end_s, m.source) == (base.peak_s, base.start_s, base.end_s, "chat")
|
|
assert m.reasons["face"]["label"] == "risada" and m.reasons["face"]["onsetS"] == 296
|
|
assert 0 < m.score - base.score <= Params().face_bonus_max + 1e-9
|
|
|
|
|
|
def test_face_alone_never_becomes_a_clip():
|
|
# Risada no rosto sem áudio nem chat se mexendo: não vira corte (o rosto reforça, não manda).
|
|
face = laugh_at(calm_face(600), 300)
|
|
assert find_moments(600, [], noisy_loudness(600), [], face=face) == []
|
|
|
|
|
|
def test_face_with_audio_support_creates_face_moment():
|
|
loud = noisy_loudness(600)
|
|
for t in range(299, 305):
|
|
loud[t] = -28.5 # um pouco acima do normal: sozinho não chega a pico de áudio
|
|
assert find_moments(600, [], loud, []) == []
|
|
moments = find_moments(600, [], loud, [], face=laugh_at(calm_face(600), 300))
|
|
assert len(moments) == 1
|
|
m = moments[0]
|
|
assert m.source == "face" and 300 <= m.peak_s <= 305 and m.reasons["face"]["label"] == "risada"
|
|
assert m.start_s < 300 - 5 and m.reasons["support"]["audioZ"] >= 1.0 and m.score > 0
|
|
|
|
|
|
def test_face_moment_respects_live_settle():
|
|
loud = noisy_loudness(316)
|
|
for t in range(299, 305):
|
|
loud[t] = -28.5
|
|
assert find_moments(316, [], loud, [], live=True, face=laugh_at(calm_face(316), 300)) == []
|