import re from .common import InfoExtractor, ExtractorError from ..utils import ( clean_html, determine_ext, int_or_none, float_or_none, js_to_json, mimetype2ext, traverse_obj, urljoin, url_or_none, ) class RTVCPlayBaseIE(InfoExtractor): _BASE_VALID_URL = r'https?://(?:www\.)?rtvcplay\.co' def _extract_player_config(self, webpage, video_id): return self._search_json( r']*>[^<]*(?:var|let|const)\s+config\s*=', re.sub(r'"\s*\+\s*"', '', webpage), 'player_config', video_id, transform_source=js_to_json) def _extract_formats_and_subtitles_player_config(self, player_config, video_id): formats, subtitles = [], {} for source in traverse_obj(player_config, ('sources', ..., lambda _, v: url_or_none(v['url']))): ext = mimetype2ext(source.get('mimetype'), default=determine_ext(source['url'])) if ext == 'm3u8': fmts, subs = self._extract_m3u8_formats_and_subtitles( source['url'], video_id, 'mp4', fatal=False) formats.extend(fmts) self._merge_subtitles(subs, target=subtitles) else: formats.append({ 'url': source['url'], 'ext': ext, }) return formats, subtitles class RTVCPlayIE(RTVCPlayBaseIE): _VALID_URL = RTVCPlayBaseIE._BASE_VALID_URL + r'/(?P(?!embed)[^/]+)/(?:[^?#]+/)?(?P[\w-]+)' _TESTS = [{ 'url': 'https://www.rtvcplay.co/en-vivo/canal-institucional', 'info_dict': { 'id': 'canal-institucional', 'title': r're:^Canal Institucional', 'description': 'md5:eff9e548394175928059320c006031ea', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', 'live_status': 'is_live', 'ext': 'mp4', }, 'params': { 'skip_download': 'Livestream', }, }, { 'url': 'https://www.rtvcplay.co/en-vivo/senal-colombia', 'info_dict': { 'id': 'senal-colombia', 'title': r're:^Señal Colombia', 'description': 'md5:799f16a401d97f40c33a2c6a3e2a507b', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', 'live_status': 'is_live', 'ext': 'mp4', }, 'params': { 'skip_download': 'Livestream', }, }, { 'url': 'https://www.rtvcplay.co/en-vivo/radio-nacional', 'info_dict': { 'id': 'radio-nacional', 'title': r're:^Radio Nacional', 'description': 'md5:5de009bc6a9fa79d2a6cf0b73f977d53', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', 'live_status': 'is_live', 'ext': 'mp4', }, 'params': { 'skip_download': 'Livestream', }, }, { 'url': 'https://www.rtvcplay.co/peliculas-ficcion/senoritas', 'md5': '1288ee6f6d1330d880f98bff2ed710a3', 'info_dict': { 'id': 'senoritas', 'title': 'Señoritas', 'description': 'md5:f095a2bb52cb6cf279daf6302f86fb32', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', 'ext': 'mp4', }, }, { 'url': 'https://www.rtvcplay.co/competencias-basicas-ciudadanas-y-socioemocionales/profe-en-tu-casa/james-regresa-clases-28022022', 'md5': 'f040a7380a269ad633cf837384d5e9fc', 'info_dict': { 'id': 'james-regresa-clases-28022022', 'title': 'James regresa a clases - 28/02/2022', 'description': 'md5:c5dcdf757c7ab29305e8763c6007e675', 'ext': 'mp4', }, }, { 'url': 'https://www.rtvcplay.co/peliculas-documentales/llinas-el-cerebro-y-el-universo', 'info_dict': { 'id': 'llinas-el-cerebro-y-el-universo', 'title': 'Llinás, el cerebro y el universo', 'description': 'md5:add875bf2309bb52b3e8b9b06116d9b0', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', }, 'playlist_mincount': 3, }, { 'url': 'https://www.rtvcplay.co/competencias-basicas-ciudadanas-y-socioemocionales/profe-en-tu-casa', 'info_dict': { 'id': 'profe-en-tu-casa', 'title': 'Profe en tu casa', 'description': 'md5:47dbe20e263194413b1db2a2805a4f2e', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', }, 'playlist_mincount': 537, }, { 'url': 'https://www.rtvcplay.co/series-al-oido/relato-de-un-naufrago-una-travesia-del-periodismo-a-la-literatura', 'info_dict': { 'id': 'relato-de-un-naufrago-una-travesia-del-periodismo-a-la-literatura', 'title': 'Relato de un náufrago: una travesía del periodismo a la literatura', 'description': 'md5:6da28fdca4a5a568ea47ef65ef775603', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', }, 'playlist_mincount': 5, }, { 'url': 'https://www.rtvcplay.co/series-al-oido/diez-versiones', 'info_dict': { 'id': 'diez-versiones', 'title': 'Diez versiones', 'description': 'md5:997471ed971cb3fd8e41969457675306', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', }, 'playlist_mincount': 20, }] def _real_extract(self, url): video_id, category = self._match_valid_url(url).group('id', 'category') webpage = self._download_webpage(url, video_id) hydration = self._search_json( r'window\.__RTVCPLAY_STATE__\s*=', webpage, 'hydration', video_id, transform_source=js_to_json)['content']['currentContent'] asset_id = traverse_obj(hydration, ('video', 'assetid')) if asset_id: hls_url = hydration['base_url_hls'].replace('[node:field_asset_id]', asset_id) else: hls_url = traverse_obj(hydration, ('channel', 'hls')) metadata = traverse_obj(hydration, { 'title': 'title', 'description': 'description', 'thumbnail': ((('channel', 'image', 'logo'), ('resource', 'image', 'cover_desktop')), 'path'), }, get_all=False) # Probably it's a program's page if not hls_url: seasons = traverse_obj( hydration, ('widgets', lambda _, y: y['type'] == 'seasonList', 'contents'), get_all=False) if not seasons: podcast_episodes = hydration.get('audios') if not podcast_episodes: raise ExtractorError('Could not find asset_id nor program playlist nor podcast episodes') return self.playlist_result([ self.url_result(episode['file'], url_transparent=True, **traverse_obj(episode, { 'title': 'title', 'description': ('description', {clean_html}), 'episode_number': ('chapter_number', {float_or_none}, {int_or_none}), 'season_number': ('season', {int_or_none}), })) for episode in podcast_episodes], video_id, **metadata) entries = [self.url_result( urljoin(url, episode['slug']), url_transparent=True, **traverse_obj(season, { 'season': 'title', 'season_number': ('season', {int_or_none}), }), **traverse_obj(episode, { 'title': 'title', 'thumbnail': ('image', 'cover', 'path'), 'episode_number': ('chapter_number', {int_or_none}), })) for season in seasons for episode in traverse_obj(season, ('contents', ...))] return self.playlist_result(entries, video_id, **metadata) formats, subtitles = self._extract_m3u8_formats_and_subtitles(hls_url, video_id, 'mp4') return { 'id': video_id, 'formats': formats, 'subtitles': subtitles, 'is_live': category == 'en-vivo', **metadata, } class RTVCPlayEmbedIE(RTVCPlayBaseIE): _VALID_URL = RTVCPlayBaseIE._BASE_VALID_URL + r'/embed/(?P[\w-]+)' _TESTS = [{ 'url': 'https://www.rtvcplay.co/embed/72b0e699-248b-4929-a4a8-3782702fa7f9', 'md5': 'ed529aeaee7aa2a72afe91ac7d1177a8', 'info_dict': { 'id': '72b0e699-248b-4929-a4a8-3782702fa7f9', 'title': 'Tráiler: Señoritas', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', 'ext': 'mp4', } }] def _real_extract(self, url): video_id = self._match_id(url) webpage = self._download_webpage(url, video_id) player_config = self._extract_player_config(webpage, video_id) formats, subtitles = self._extract_formats_and_subtitles_player_config(player_config, video_id) asset_id = traverse_obj(player_config, ('rtvcplay', 'assetid')) metadata = {} if not asset_id else self._download_json( f'https://cms.rtvcplay.co/api/v1/video/asset-id/{asset_id}', video_id, fatal=False) return { 'id': video_id, 'formats': formats, 'subtitles': subtitles, **traverse_obj(metadata, { 'title': 'title', 'description': 'description', 'thumbnail': ('image', ..., 'thumbnail', 'path'), }, get_all=False) } class RTVCKalturaIE(RTVCPlayBaseIE): _VALID_URL = r'https?://media\.rtvc\.gov\.co/kalturartvc/(?P[\w-]+)' _TESTS = [{ 'url': 'https://media.rtvc.gov.co/kalturartvc/indexSC.html', 'info_dict': { 'id': 'indexSC', 'title': r're:^Señal Colombia', 'description': 'md5:799f16a401d97f40c33a2c6a3e2a507b', 'thumbnail': r're:^https?://.*\.(?:jpg|png)', 'live_status': 'is_live', 'ext': 'mp4', }, 'params': { 'skip_download': 'Livestream', }, }] def _real_extract(self, url): video_id = self._match_id(url) webpage = self._download_webpage(url, video_id) player_config = self._extract_player_config(webpage, video_id) formats, subtitles = self._extract_formats_and_subtitles_player_config(player_config, video_id) channel_id = traverse_obj(player_config, ('rtvcplay', 'channelId')) metadata = {} if not channel_id else self._download_json( f'https://cms.rtvcplay.co/api/v1/taxonomy_term/streaming/{channel_id}', video_id, fatal=False) fmts, subs = self._extract_m3u8_formats_and_subtitles( traverse_obj(metadata, ('channel', 'hls')), video_id, 'mp4', fatal=False) formats.extend(fmts) self._merge_subtitles(subs, target=subtitles) return { 'id': video_id, 'formats': formats, 'subtitles': subtitles, 'is_live': True, **traverse_obj(metadata, { 'title': 'title', 'description': 'description', 'thumbnail': ('channel', 'image', 'logo', 'path'), }) }