Clipador: fontes em inglês (juízes em inglês, painel em português) e monitor de uso da IA

Clipador em inglês (doc/clipador/IDIOMAS.md):
- Fonte en-US: analista, editor e varredura com prompts, glossário de chat e saída em inglês;
  a mesma resposta traz o bloco `pt` (tradução para o painel, sem chamada extra).
- Fala do corte traduzida por frase (clips.translation_json, migração v13) e
  POST /v1/clips/{id}/translate.
- caption.translation (só para desenvolver): português embaixo da legenda em inglês.
- Risada e gírias do chat em inglês na heurística (LMAO, LOL, 💀, 😭, no way, bruh...).
- Idioma do público → código do Whisper (en-US dava 422 no serviço de voz).
- Painel: bloco "Em português" no corte, análise/resumo em PT, título traduzido na lista,
  coluna "Ao vivo" mostra a gravação ativa da fonte.

IA: monitor de uso do limite (ai-usage) e 4º provedor (Antigravity CLI).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-11-18 03:07:09 -03:00
co-authored by Claude Opus 5.5
parent ed31d52df0
commit 20a9805d2a
52 changed files with 1808 additions and 311 deletions
+42
View File
@@ -311,3 +311,45 @@ def test_api_accepts_gemini(env): # noqa: F811
body = r.json()
assert body["provider"] == "gemini" and body["model"] == "gemini-3.5-flash" and "gemini" in body["models"]
assert "chain" in body["fallback"] and isinstance(body["pauses"], list)
# ---------------------------------------------------------------- uso e limites (data/ai/usage.json do painel)
def write_usage(folder: Path, providers: dict, limit: int = 90, age_s: float = 0) -> None:
import time as _time
from datetime import datetime as _dt, timezone as _tz
iso = lambda t: _dt.fromtimestamp(t, _tz.utc).isoformat().replace("+00:00", "Z") # noqa: E731
now = _time.time()
data = {"updatedAt": iso(now), "limitPercent": limit, "providers": {
name: {"at": iso(now - age_s), "source": "cli", "blocked": spec.get("blocked", False), "plan": None, "note": None,
"error": None, "windows": [{"id": "five_hour", "label": "Sessão (5 h)", "usedPercent": spec["used"],
"resetsAt": iso(now + spec.get("resets_in", 3600))}]}
for name, spec in providers.items()}}
folder.mkdir(parents=True, exist_ok=True)
(folder / "usage.json").write_text(json.dumps(data), encoding="utf-8")
def test_over_limit_rule(ai_dir):
ai = AiConfig(ai_dir, cipher=ReversedCipher(), environ={})
assert ai.over_limit("claude") is None # sem arquivo
write_usage(ai_dir, {"claude": {"used": 93}, "codex": {"used": 100, "blocked": True}, "gemini": {"used": 50}})
assert ai.over_limit("claude") == "Sessão (5 h) em 93% (limite 90%)"
assert ai.over_limit("codex") == "bateu o limite (Sessão (5 h))"
assert ai.over_limit("gemini") is None
write_usage(ai_dir, {"claude": {"used": 93}}, limit=95)
assert ai.over_limit("claude") is None
write_usage(ai_dir, {"codex": {"used": 100, "blocked": True, "resets_in": -60}})
assert ai.over_limit("codex") is None # a janela já zerou
write_usage(ai_dir, {"claude": {"used": 99}}, age_s=3600)
assert ai.over_limit("claude") is None # leitura velha não decide
(ai_dir / "usage.json").write_text("{quebrado", encoding="utf-8")
assert ai.over_limit("claude") is None
def test_chain_puts_provider_near_limit_after_free_ones(make_judge, ai_dir):
judge = make_judge(provider="codex")
write_usage(ai_dir, {"codex": {"used": 100, "blocked": True}})
assert judge.fallback_chain() == ["claude", "codex"] # escolhida no limite: a livre vem antes
write_usage(ai_dir, {"codex": {"used": 95}, "claude": {"used": 97}})
assert judge.fallback_chain() == ["codex", "claude"] # todas no limite: melhor tentar do que parar
write_settings(ai_dir, {**settings_with(), "fallback": {"enabled": False, "cooldownMin": 15}})
assert judge.fallback_chain() == ["codex"] # reserva desligada: só a escolhida, mesmo no limite
+150
View File
@@ -0,0 +1,150 @@
"""Fonte em inglês (18/11/2026, doc/clipador/IDIOMAS.md): juízes em inglês com o português do painel na mesma
resposta, risada/gíria do chat em inglês, fala do corte traduzida por frase e legenda em português de desenvolvimento
(desligada por padrão)."""
import json
from pathlib import Path
from app import editor, editor_render, layout
from app.judge import (ANALYSIS_SCHEMA_EN, JUDGE_SYSTEM_EN, SCAN_SCHEMA_EN, VERDICT_SCHEMA_EN, Judge, Verdict,
analysis_block, build_prompt, lang_key)
from app.scoring import REACTION_PATTERNS, ChatMessage, Word
from app.subtitles import phrases
from .test_editor import validate
from .test_render_and_duplicates import add_clip
from .test_sources_and_api import env, seed_session # noqa: F401 - fixture reaproveitada
class EnglishJudge(Judge):
"""Juiz falso: registra o idioma pedido e devolve o bloco `pt`, como os schemas em inglês exigem."""
def __init__(self):
super().__init__("codex", "fake", Path("."))
self.seen: dict = {}
self.translated: list[list[str]] = []
def analyze(self, prompt, images, lang="pt"):
self.seen["analyze"] = (lang, prompt)
return {"situation": "jogada", "what_happened": "He robbed the bank.", "screen": "", "speech": "",
"chat_meaning": "", "confidence": 0.9, "action_start_s": 0, "reaction_end_s": 0, "hook": "NO WAY",
"pt": {"what_happened": "Ele roubou o banco.", "screen": "", "speech": "", "chat_meaning": "", "hook": "NÃO ACREDITO"}}
def judge(self, prompt, start_s, end_s, images=None, lang="pt", **_limits):
self.seen["judge"] = (lang, prompt)
return Verdict(True, 8.0, start_s, end_s, "He robbed the bank solo", "Insane solo heist.", "great play", "fake",
("heist",), pt={"title": "Ele roubou o banco sozinho", "caption": "Assalto solo insano.",
"reason": "jogada ótima"})
def translate_lines(self, lines):
self.translated.append(list(lines))
return [f"PT: {text}" for text in lines]
def _english_session(engine) -> int:
sid = seed_session(engine)
engine.db.run("UPDATE sessions SET language = 'en-US' WHERE id = ?", sid)
words = [[281.0, 281.3, "No"], [281.3, 281.6, "way"], [281.6, 282.0, "bro."], [284.0, 284.4, "Let's"], [284.4, 284.9, "go!"]]
engine.db.run("UPDATE segments SET words_json = '[]' WHERE session_id = ?", sid)
engine.db.run("UPDATE segments SET words_json = ? WHERE session_id = ? AND idx = 9", json.dumps(words), sid)
return sid
def test_lang_key():
assert lang_key("en-US") == "en" and lang_key("en") == "en"
assert lang_key("pt-BR") == "pt" and lang_key("es-419") == "pt" and lang_key("") == "pt" and lang_key(None) == "pt"
def test_english_schemas_require_the_portuguese_block():
assert ANALYSIS_SCHEMA_EN["required"][-1] == "pt" and VERDICT_SCHEMA_EN["required"][-1] == "pt"
assert "summary_pt" in SCAN_SCHEMA_EN["required"]
assert "pt" in SCAN_SCHEMA_EN["properties"]["moments"]["items"]["required"]
# Os CLIs exigem todas as propriedades em `required` (strict).
for schema in (ANALYSIS_SCHEMA_EN, VERDICT_SCHEMA_EN, SCAN_SCHEMA_EN):
assert set(schema["properties"]) == set(schema["required"])
assert "LMAO" in JUDGE_SYSTEM_EN and "💀" in JUDGE_SYSTEM_EN
def test_english_prompt_labels_and_analysis_block_without_pt():
prompt = build_prompt("Tfue", "twitch", 100, 160, 150, [Word(140, 141, "bro")], [ChatMessage(151, "LMAO")], {},
"GTA 6 Waiting Room", lang="en")
assert "Channel: Tfue" in prompt and "Transcript:" in prompt and "Chat sample:" in prompt
assert "Transcrição" not in prompt and "Canal:" not in prompt
block = analysis_block({"situation": "jogada", "pt": {"what_happened": "x"}}, "en")
assert block.startswith("ANALYST's analysis") and '"pt"' not in block
def test_english_chat_laughter_and_slang_count_as_reaction():
def tags(text):
return {tag for tag, (pattern, _) in REACTION_PATTERNS.items() if pattern.search(text)}
assert "risada" in tags("LMAO") and "risada" in tags("lol") and "risada" in tags("💀💀") and "risada" in tags("im dead")
assert "risada" in tags("😭😭") and "risada" not in tags("lollipop")
assert "susto" in tags("no way") and "susto" in tags("bruhhh") and "susto" in tags("nahhh")
assert "hype" in tags("LETS GOOO") and "hype" in tags("sheesh") and "hype" in tags("clutch")
assert "clip" in tags("clip that")
def test_phrases_split_on_sentence_end_and_pause():
words = [Word(0, 0.3, "No"), Word(0.3, 0.6, "way"), Word(0.6, 0.9, "bro."), Word(0.9, 1.2, "This"), Word(1.2, 1.5, "is"),
Word(1.5, 1.8, "crazy"), Word(1.8, 2.1, "man."), Word(3.5, 3.8, "Okay")]
assert [t for _, _, t in phrases(words)] == ["No way bro. This is crazy man.", "Okay"]
assert phrases([]) == []
def test_english_session_judges_in_english_and_stores_portuguese(env): # noqa: F811
client, engine = env
engine.judge = EnglishJudge()
sid = _english_session(engine)
add_clip(engine, sid, 1, "candidate", 285, 270, 340)
engine.judge_clip(engine.db.get("SELECT * FROM clips WHERE id = 1"))
assert engine.judge.seen["analyze"][0] == "en" and engine.judge.seen["judge"][0] == "en"
assert "Transcript:" in engine.judge.seen["judge"][1] and "ANALYST's analysis" in engine.judge.seen["judge"][1]
clip = client.get("/v1/clips/1", headers=_auth()).json()
assert clip["title"] == "He robbed the bank solo" # o vídeo/publicação: inglês
assert clip["judge"]["pt"]["title"] == "Ele roubou o banco sozinho" # o painel: português
assert clip["judge"]["analysis"]["pt"]["what_happened"] == "Ele roubou o banco."
lines = clip["translation"]["lines"]
assert [x["text"] for x in lines] == ["No way bro.", "Let's go!"]
assert lines[0]["pt"] == "PT: No way bro." and clip["translation"]["language"] == "pt-BR"
# Tradução que já cobre o trecho é reaproveitada (sem chamar a IA de novo).
engine.translate_clip(1)
assert len(engine.judge.translated) == 1
def test_portuguese_session_has_no_translation(env): # noqa: F811
client, engine = env
engine.judge = EnglishJudge()
sid = seed_session(engine)
add_clip(engine, sid, 1, "candidate", 300, 280, 350)
engine.judge_clip(engine.db.get("SELECT * FROM clips WHERE id = 1"))
assert engine.judge.seen["judge"][0] == "pt" and "Transcrição:" in engine.judge.seen["judge"][1]
assert client.get("/v1/clips/1", headers=_auth()).json()["translation"] is None
assert client.post("/v1/clips/1/translate", headers=_auth()).status_code == 409
def test_dev_translation_line_only_when_enabled(tmp_path):
clip = {"session_id": 1, "start_s": 10.0, "end_s": 75.0, "peak_s": 50.0, "title": "He robbed the bank", "source": "chat"}
words = [Word(20, 20.4, "No"), Word(20.4, 20.8, "way"), Word(20.8, 21.2, "bro.")]
lines = [{"s": 20.0, "e": 21.2, "text": "No way bro.", "pt": "Não acredito, mano."}]
for enabled in (False, True):
cfg = layout.merge({"caption": {"translation": enabled}})
p = validate(editor.auto_project(clip, None, words, [], 100.0, cfg), tmp_path)
ass = editor_render.build_project_ass(p, editor.effective_settings(cfg, p), "twitch.tv/tfue", translation=lines)
has = "Não acredito, mano." in ass and "Style: Trans," in ass
assert has is enabled
# Sem tradução disponível nada muda, mesmo ligado.
cfg = layout.merge({"caption": {"translation": True}})
p = validate(editor.auto_project(clip, None, words, [], 100.0, cfg), tmp_path)
assert "Style: Trans," not in editor_render.build_project_ass(p, editor.effective_settings(cfg, p), "x")
assert layout.DEFAULTS["caption"]["translation"] is False
assert layout.validate({"caption": {"translation": True}})["caption"]["translation"] is True
def _auth():
from .test_sources_and_api import auth
return auth()
def test_voice_language_maps_audience_codes_to_whisper_codes():
from app.worker import voice_language
assert voice_language("en-US") == "en" and voice_language("es-419") == "es" and voice_language("pt-BR") == "pt-BR"
assert voice_language("en-GB") == "en" and voice_language("") == "pt-BR"
+1 -1
View File
@@ -114,7 +114,7 @@ class ScanJudge(Judge):
super().__init__("codex", "fake", __import__("pathlib").Path("."))
self.prompts = []
def scan(self, prompt, images):
def scan(self, prompt, images, lang="pt"):
self.prompts.append(prompt)
return {"screen_game": "ROB RUN", "has_webcam": True, "summary": "tiroteio na base",
"moments": [{"score": 8, "start_s": 100, "end_s": 140, "title": "Emboscada", "caption": "c #robrun",
@@ -138,13 +138,19 @@ def test_language_on_channel_and_session(ready):
def test_language_note_reaches_the_ai_context(ready):
client, engine = ready
assert output_language_note("pt-BR") == "" and output_language_note("") == "" and output_language_note(None) == ""
assert "English (US)" in output_language_note("en-US")
# Inglês (18/11/2026): os juízes trabalham em inglês (prompts próprios), sem a nota; espanhol mantém a nota.
assert output_language_note("en-US") == ""
assert "Español" in output_language_note("es-419")
client.post("/v1/channels", json={"url": URL, "language": "en-US"}, headers=auth())
sid = client.post("/v1/sessions", json={"url": URL}, headers=auth()).json()["id"]
gringa = engine.db.get("SELECT * FROM sessions WHERE id = ?", sid)
assert "Idioma de saída: English (US)" in engine.intel.context(gringa)
assert engine.intel.lang(gringa) == "en"
assert "Stream title:" in engine.intel.context(gringa)
engine.db.run("UPDATE sessions SET language = 'es-419' WHERE id = ?", sid)
assert "Idioma de saída: Español" in engine.intel.context(engine.db.get("SELECT * FROM sessions WHERE id = ?", sid))
engine.db.run("UPDATE sessions SET language = '' WHERE id = ?", sid)
assert "Idioma de saída" not in engine.intel.context(engine.db.get("SELECT * FROM sessions WHERE id = ?", sid))
br = engine.db.get("SELECT * FROM sessions WHERE id = ?", sid)
assert "Idioma de saída" not in engine.intel.context(br) and "Título da live:" in engine.intel.context(br)
def test_transcription_uses_session_language(ready, monkeypatch):
+1 -1
View File
@@ -20,7 +20,7 @@ class RecuoJudge(Judge):
super().__init__("codex", "fake", Path("."))
self.back_to = back_to
def analyze(self, prompt, images):
def analyze(self, prompt, images, lang="pt"):
return {"situation": "lag_ou_bug", "confidence": 0.9, "action_start_s": 0, "reaction_end_s": 0, "hook": ""}
def judge(self, prompt, start_s, end_s, images=None, **_limits):