utilitas 2001.1.136 → 2001.1.139

package/lib/alan.mjs

@@ -8,9 +8,9 @@ import {
     STREAM, FILE, BASE64, BUFFER, DATAURL, MIME_TEXT, MIME_PNG, MIME_JPEG,
     MIME_MOV, MIME_MPEG, MIME_MP4, MIME_MPG, MIME_AVI, MIME_WMV, MIME_MPEGPS,
     MIME_FLV, MIME_GIF, MIME_WEBP, MIME_PDF, MIME_AAC, MIME_FLAC, MIME_MP3,
-    MIME_MPEGA, MIME_M4A, MIME_MPGA, MIME_OPUS, MIME_PCM, MIME_WAV, MIME_WEBM,
-    MIME_TGPP, MIME_PCM16, MIME_OGG, convert, decodeBase64DataURL, getImageSize,
-    getMime, getTempPath,
+    MIME_MPEGA, MIME_M4A, MIME_MPGA, MIME_OPUS, MIME_PCM, MIME_WAV, MIME_WAVE,
+    MIME_WEBM, MIME_TGPP, MIME_PCM16, MIME_OGG, convert, decodeBase64DataURL,
+    getImageSize, getMime, getTempPath,
 } from './storage.mjs';
 
 import {
@@ -45,37 +45,32 @@ const STT_PROMPT = 'Please transcribe the audio into clean text. Return only the
 const _NEED = ['OpenAI', '@google/genai'];
 
 const [
-    OPENAI, GOOGLE, OLLAMA, NOVA, DEEPSEEK_32, MD_CODE, CLOUD_OPUS_46, AUDIO,
-    WAV, OPENAI_VOICE, GPT_REASONING_EFFORT, THINK, THINK_STR, THINK_END,
-    TOOLS_STR, TOOLS_END, TOOLS, TEXT, OK, FUNC, GPT_54, GPT_IMAGE_15,
-    GEMMA_3_27B, ANTHROPIC, ais, MAX_TOOL_RECURSION, LOG, name, user, system,
-    assistant, JSON_OBJECT, PROMPT_IS_REQUIRED, k, trimTailing, trimBeginning,
-    GEMINI_30_PRO_IMAGE, IMAGE, JINA, SILICONFLOW, SF_DEEPSEEK_32,
-    OPENROUTER_API, OPENROUTER, AUTO, TOOL, ONLINE, GEMINI_31_PRO,
-    GEMINI_30_FLASH, VEO_31, IMAGEN_4_UPSCALE, ERROR_GENERATING,
-    GEMINI_25_FLASH_TTS, GEMINI_25_PRO_TTS, wav, PNG_EXT, GPT_4O_MIMI_TTS,
-    GPT_4O_TRANSCRIBE, INVALID_AUDIO, OGG_EXT, ELLIPSIS, TOP_LIMIT, ATTACHMENT,
-    PROCESSING, CURSOR, LN1, LN2, TOP, DEEPSEEK, DEEP_RESEARCH_PRO,
-    MD_CODE_ESCAPED, VERTEX, LYRIA_002,
+    OPENAI, GOOGLE, OLLAMA, NOVA, MD_CODE, CLOUD_OPUS_47, AUDIO, WAV,
+    OPENAI_VOICE, GPT_REASONING_EFFORT, THINK, THINK_STR, THINK_END, TOOLS_STR,
+    TOOLS_END, TOOLS, TEXT, OK, FUNC, GPT_55, GPT_IMAGE_2, GEMMA_4_31B,
+    ANTHROPIC, ais, MAX_TOOL_RECURSION, LOG, name, user, system, assistant,
+    JSON_OBJECT, PROMPT_IS_REQUIRED, k, trimTailing, trimBeginning,
+    GEMINI_30_PRO_IMAGE, IMAGE, OPENROUTER_API, OPENROUTER, AUTO, TOOL, ONLINE,
+    GEMINI_31_PRO, GEMINI_30_FLASH, VEO_31, IMAGEN_4_UPSCALE, ERROR_GENERATING,
+    GEMINI_25_PRO_TTS, wav, GPT_AUDIO, ELLIPSIS, TOP_LIMIT, ATTACHMENT,
+    PROCESSING, CURSOR, LN1, LN2, TOP, DEEP_RESEARCH_MAX, MD_CODE_ESCAPED,
+    VERTEX, LYRIA_3, GPT_55_MINI,
 ] = [
-        'OpenAI', 'Google', 'Ollama', 'nova', 'deepseek-3.2-speciale', '```',
-        'claude-opus-4.6', 'audio', 'wav', 'OPENAI_VOICE', 'medium', 'think',
-        '<think>', '</think>', '<tools>', '</tools>', 'tools', 'text', 'OK',
-        'function', 'gpt-5.4', 'gpt-image-1.5', 'gemma3:27b', 'Anthropic', [],
-        30, { log: true }, 'Alan', 'user', { role: 'system' },
-        { role: 'assistant' }, 'json_object', 'Prompt is required.',
-        x => 1000 * x, x => x.replace(/[\.\s]*$/, ''),
-        x => x.replace(/^[\.\s]*/, ''), 'gemini-3-pro-image-preview', 'image',
-        'Jina', 'SiliconFlow', 'deepseek-ai/DeepSeek-V3.2-exp',
-        'https://openrouter.ai/api/v1', 'OpenRouter', 'openrouter/auto', 'tool',
-        ':online', 'gemini-3.1-pro-preview', 'gemini-3-flash-preview',
+        'OpenAI', 'Google', 'Ollama', 'nova', '```', 'claude-opus-4.7', 'audio',
+        'wav', 'OPENAI_VOICE', 'medium', 'think', '<think>', '</think>',
+        '<tools>', '</tools>', 'tools', 'text', 'OK', 'function', 'gpt-5.5',
+        'gpt-5.4-image-2', 'gemma4:31b', 'Anthropic', [], 30, { log: true },
+        'Alan', 'user', { role: 'system' }, { role: 'assistant' },
+        'json_object', 'Prompt is required.', x => 1000 * x,
+        x => x.replace(/[\.\s]*$/, ''), x => x.replace(/^[\.\s]*/, ''),
+        'gemini-3-pro-image-preview', 'image', 'https://openrouter.ai/api/v1',
+        'OpenRouter', 'openrouter/auto', 'tool', ':online',
+        'gemini-3.1-pro-preview', 'gemini-3-flash-preview',
         'veo-3.1-generate-preview', 'imagen-4.0-upscale-preview',
-        'Error generating content.', 'gemini-2.5-flash-preview-tts',
-        'gemini-2.5-pro-preview-tts', 'wav', 'png', 'gpt-4o-mini-tts',
-        'gpt-4o-transcribe', 'Invalid audio data.', 'ogg', '...', 3,
-        'ATTACHMENT', { processing: true }, ' █', '\n\n', '\n\n\n', 'top',
-        'DeepSeek', 'deep-research-pro-preview-12-2025', '\`\`\`', 'Vertex',
-        'lyria-002',
+        'Error generating content.', 'gemini-2.5-pro-preview-tts', 'wav',
+        'gpt-audio', '...', 3, 'ATTACHMENT', { processing: true }, ' █', '\n\n',
+        '\n\n\n', 'top', 'deep-research-max-preview-04-2026', '\`\`\`',
+        'Vertex', 'lyria-3-pro-preview', 'gpt-5-mini',
     ];
 
 const [joinL1, joinL2]
@@ -98,70 +93,41 @@ const m = x => k(k(x));
 const [MAX_TOKENS, ATTACHMENT_TOKEN_COST] = [m(1), k(10)];
 
 const MODEL_ICONS = {
-    [OPENROUTER]: '🔀', [OPENAI]: '⚛️', [GOOGLE]: '♊️', // [JINA]: '✴️',
-    [OLLAMA]: '🦙', [ANTHROPIC]: '✳️', [SILICONFLOW]: '🧬', [DEEPSEEK]: '🐬',
-    [VERTEX]: '📐',
+    [OPENROUTER]: '🔀', [OPENAI]: '⚛️', [GOOGLE]: '♊️', [OLLAMA]: '🦙',
+    [ANTHROPIC]: '✳️', [VERTEX]: '❇️',
 };
 
 const FEATURE_ICONS = {
     audio: '🔊', 'deep-research': '🔍', fast: '⚡️', hearing: '👂', hidden: '🙈',
-    image: '🎨', reasoning: '🧠', structured: '📊', tools: '🧰', video: '🎬',
-    vision: '👁️', // finetune: '🔧',
-};
-
-const DEEPSEEK_32_RULES = {
-    source: DEEPSEEK, contextWindow: k(163.8), maxOutputTokens: k(65.5),
-    structured: true, tools: true, reasoning: true,
+    image: '🎨', music: '🎵', reasoning: '🧠', structured: '📊', tools: '🧰',
+    video: '🎬', vision: '👁️', // finetune: '🔧',
 };
 
 const MODELS = {
     // models with generation capabilities
-    [LYRIA_002]: {
-        source: VERTEX, maxInputTokens: 480,
-        audio: true, defaultProvider: VERTEX,
-    },
+    [LYRIA_3]: { music: true },
     [IMAGEN_4_UPSCALE]: {
         source: GOOGLE, maxInputTokens: 0, image: true, vision: true,
         supportedMimeTypes: [MIME_PNG, MIME_JPEG], defaultProvider: VERTEX,
     },
     [VEO_31]: {
-        source: GOOGLE, maxInputTokens: 1024, attachmentTokenCost: 0,
+        source: GOOGLE, maxInputTokens: k(1), attachmentTokenCost: 0,
         video: true, vision: true,
         supportedMimeTypes: [MIME_PNG, MIME_JPEG], defaultProvider: GOOGLE,
     },
-    [GPT_IMAGE_15]: {
-        source: OPENAI, contextWindow: k(400), maxOutputTokens: k(128),
-        hearing: true, image: true, reasoning: true, structured: true,
-        tools: true, vision: true, defaultProvider: OPENAI,
-        supportedMimeTypes: [MIME_PNG, MIME_JPEG, MIME_GIF, MIME_WEBP],
-    },
     // tts/stt models
-    [GEMINI_25_FLASH_TTS]: {
-        source: GOOGLE, maxInputTokens: k(32),
-        audio: true, fast: true, hidden: true, defaultProvider: GOOGLE,
-    },
     [GEMINI_25_PRO_TTS]: {
         source: GOOGLE, maxInputTokens: k(32),
         audio: true, hidden: true, defaultProvider: GOOGLE,
     },
-    [GPT_4O_MIMI_TTS]: {
-        source: OPENAI, maxInputTokens: k(2),
-        audio: true, fast: true, hidden: true, defaultProvider: OPENAI,
-    },
-    [GPT_4O_TRANSCRIBE]: {
-        source: OPENAI, maxInputTokens: 0,
-        hearing: true, fast: true, hidden: true, defaultProvider: OPENAI,
-    },
+    [GPT_AUDIO]: { hidden: true, },
     // agents with deep-research capabilities
-    [DEEP_RESEARCH_PRO]: { // https://ai.google.dev/gemini-api/docs/deep-research
+    [DEEP_RESEARCH_MAX]: { // https://ai.google.dev/gemini-api/docs/deep-research
         source: GOOGLE, contextWindow: m(1.05), maxOutputTokens: k(65.5),
         'deep-research': true, reasoning: true, defaultProvider: GOOGLE,
     },
-    // best Chinese models
-    [DEEPSEEK_32]: DEEPSEEK_32_RULES,
-    [SF_DEEPSEEK_32]: { ...DEEPSEEK_32_RULES, defaultProvider: SILICONFLOW },
     // best local model
-    [GEMMA_3_27B]: {
+    [GEMMA_4_31B]: {
         source: GOOGLE, contextWindow: k(128), maxOutputTokens: k(8),
         fast: true, structured: true, vision: true,
         supportedMimeTypes: [MIME_PNG, MIME_JPEG, MIME_GIF],
@@ -186,7 +152,7 @@ for (const n in MODELS) {
 }
 // Auto model have some issues with tools and reasoning, so we disable them here
 // MODELS[AUTO] = { name: AUTO, defaultProvider: OPENROUTER, };
-// for (const n of [GPT_54, GEMINI_31_PRO, GEMINI_30_FLASH]) {
+// for (const n of [GPT_55, GEMINI_31_PRO, GEMINI_30_FLASH]) {
 //     // get the most restrictive limits
 //     for (const key of [
 //         'contextWindow', 'maxInputTokens', 'maxOutputTokens',
@@ -224,19 +190,23 @@ for (const n in MODELS) {
 // Default models for each provider
 const DEFAULT_MODELS = {
     [OPENROUTER]: [
-        // fast and balanced models
+        // GPT: fast
+        GPT_55_MINI,
+        // GPT: strong
+        GPT_55,
+        // Gemini: fast
         GEMINI_30_FLASH,
-        // strong and fast
-        GPT_54,
-        // stronger but slow
+        // Gemini: strong
         GEMINI_31_PRO,
-        // models with generation capabilities
+        // Claude: Opus
+        CLOUD_OPUS_47,
+        // Multimedia generation models
         GEMINI_30_PRO_IMAGE,
-        // models with code capabilities
-        CLOUD_OPUS_46,
+        LYRIA_3,
+        // tts/stt models
+        // GPT_AUDIO,
     ],
-    [SILICONFLOW]: SF_DEEPSEEK_32,
-    [OLLAMA]: GEMMA_3_27B,
+    [OLLAMA]: GEMMA_4_31B,
     [OPENAI_VOICE]: NOVA,
 };
 
@@ -273,13 +243,6 @@ const adaptOpenRouterModels = async () => {
         model.source = model.label.split(':')[0] || model.id.split('/')[0];
         model.outputModalities = model.architecture.output_modalities;
         model.defaultProvider = OPENROUTER;
-        [
-            'architecture', 'canonical_slug', 'context_length', 'created',
-            'default_parameters', 'expiration_date', 'hugging_face_id',
-            'per_request_limits', 'pricing', 'supported_parameters',
-            'top_provider',
-        ].map(key => delete model[key]);
-        // Backward compatibility: {
         if (model.inputModalities.includes('image')
             || model.inputModalities.includes('video')
             || model.inputModalities.includes('file')) {
@@ -294,18 +257,15 @@ const adaptOpenRouterModels = async () => {
         if (model.outputModalities.includes('image')) {
             model.image = true;
         }
-        // if (model.outputModalities.includes('audio')) {
-        //     model.audio = true;
-        // }
+        if (model.outputModalities.includes('audio')) {
+            model.audio = true;
+        }
         // https://gemini.google.com/app/c680748b3307790b
         // issue with json output via OpenRouter
-        if (/structured/i.test(model.description)
-            && model.name !== 'gemini-3-flash') {
-            model.structured = true;
-        }
-        if (/tool/i.test(model.description)) {
-            model.tools = true;
-        }
+        model.structured = !!(model.supported_parameters?.includes('structured_outputs')
+            && ![GEMINI_30_FLASH].includes(model.name));
+        model.tools = !!(model.supported_parameters?.includes('tools')
+            || model.supported_parameters?.includes('tool_choice'));
         if (/flash|lite|mini|nano|tiny|small/i.test(model.name)) {
             model.fast = true;
         }
@@ -322,7 +282,8 @@ const adaptOpenRouterModels = async () => {
                 if (model.hearing) {
                     model.supportedMimeTypes.push(
                         MIME_AAC, MIME_FLAC, MIME_MP3, MIME_MPEGA, MIME_M4A,
-                        MIME_MPGA, MIME_OPUS, MIME_PCM, MIME_WAV, MIME_OGG,
+                        MIME_MPGA, MIME_OPUS, MIME_PCM, MIME_WAV, MIME_WAVE,
+                        MIME_OGG,
                     );
                 }
                 if (model.name === GEMINI_30_PRO_IMAGE) {
@@ -344,10 +305,7 @@ const adaptOpenRouterModels = async () => {
                     );
                 }
                 if (model.hearing) {
-                    model.supportedMimeTypes.push(MIME_WAV);
-                }
-                if (model.name === GPT_54) {
-                    model.fast = true;
+                    model.supportedMimeTypes.push(MIME_WAV, MIME_WAVE);
                 }
                 break;
             case ANTHROPIC:
@@ -359,15 +317,21 @@ const adaptOpenRouterModels = async () => {
                 }
                 break;
         }
-        // }
+        [
+            'architecture', 'canonical_slug', 'context_length', 'created',
+            'default_parameters', 'expiration_date', 'hugging_face_id',
+            'per_request_limits', 'pricing', 'top_provider',
+            'supported_parameters',
+        ].map(key => delete model[key]);
         model.label = model.label.split(':').pop().trim();
-        MODELS[model.name] = model;
+        MODELS[model.name] = { ...MODELS[model.name] || {}, ...model };
+        MODELS[model.name].music && (MODELS[model.name].audio = false);
     }
 };
 
 const unifyProvider = provider => {
     assert(provider = (provider || '').trim(), 'AI provider is required.');
-    for (let type of [OPENROUTER, GOOGLE, OPENAI, OLLAMA, SILICONFLOW, VERTEX]) { // JINA,
+    for (let type of [OPENROUTER, GOOGLE, OPENAI, OLLAMA, VERTEX]) {
         if (insensitiveCompare(provider, type)) { return type; }
     }
     throwError(`Invalid AI provider: ${provider}.`);
@@ -487,7 +451,7 @@ const setupAi = ai => {
 
 const OpenAI = async opts => {
     const lib = await libOpenAi(opts);
-    return { toFile: lib.toFile, client: new (lib).OpenAI(opts) };
+    return { client: new (lib).OpenAI(opts) };
 };
 
 const init = async (options = {}) => {
@@ -537,6 +501,10 @@ const init = async (options = {}) => {
             const location = options.location || 'global';
             var client = new GoogleGenAI({
                 vertexai: true, location, ...options,
+                googleAuthOptions: {
+                    keyFile: options.credentials,
+                    scopes: ['https://www.googleapis.com/auth/cloud-platform']
+                },
             });
             const auth = new GoogleAuth({
                 keyFile: options.credentials,
@@ -550,17 +518,6 @@ const init = async (options = {}) => {
                 });
             }
             break;
-        case OPENAI:
-            assertApiKey(provider, options);
-            var { client, toFile } = await OpenAI({ ...options });
-            for (let model of models) {
-                setupAi({
-                    provider, model, client, toFile,
-                    prompt: model.defaultProvider === OPENAI
-                        ? promptOpenAI : promptOpenRouter, priority,
-                });
-            }
-            break;
         case OLLAMA:
             // https://github.com/ollama/ollama/blob/main/docs/openai.md
             const baseURL = 'http://localhost:11434/v1/';
@@ -570,7 +527,7 @@ const init = async (options = {}) => {
             });
             for (let model of models) {
                 setupAi({
-                    provider, model, client, prompt: promptOpenRouter, priority,
+                    provider, model, client, prompt: promptOpenAI, priority,
                 });
                 ignoreErrFunc(async () => {
                     phLog(await (await fetch(`${baseURL}completions`, {
@@ -581,26 +538,16 @@ const init = async (options = {}) => {
                 }, { log: phLog });
             }
             break;
-        case SILICONFLOW:
-            assertApiKey(provider, options);
-            var client = await OpenAI({
-                baseURL: 'https://api.siliconflow.cn/v1', ...options,
-            });
-            for (let model of models) {
-                setupAi({
-                    provider, model, client, prompt: promptOpenRouter, priority,
-                });
-            }
-            break;
         default:
             assertApiKey(provider, options);
             var { client } = await OpenAI({
-                baseURL: OPENROUTER_API, ...options || {},
+                baseURL: provider === OPENROUTER ? OPENROUTER_API : undefined,
+                ...options || {},
             });
             for (let model of models) {
                 setupAi({
-                    provider: provider || OPENROUTER, model, client,
-                    prompt: promptOpenRouter, priority,
+                    provider, model, client, prompt: promptOpenAI, priority,
+                    options: { ...options || {} },
                 });
             }
     }
@@ -610,9 +557,9 @@ const init = async (options = {}) => {
 
 const packAi = (ais, options = {}) => {
     let res = options.basic ? ais.map(x => ({
-        id: x.id, name: x.name, features: x.features,
-        initOrder: x.initOrder, priority: x.priority,
-        provider: x.provider, model: x.model,
+        id: x.id, name: x.name, features: x.features, initOrder: x.initOrder,
+        priority: x.priority, provider: x.provider, model: x.model,
+        options: x.options,
     })) : ais;
     if (options.all && !Object.keys(options.select).length && !options.withHidden) {
         res = res.filter(x => !x.model.hidden);
@@ -684,7 +631,8 @@ const buildMessage = async (content, options) => {
             };
         } else if ([
             MIME_AAC, MIME_FLAC, MIME_MP3, MIME_MPEGA, MIME_M4A, MIME_MPGA,
-            MIME_OPUS, MIME_PCM, MIME_WAV, MIME_TGPP, MIME_PCM16, MIME_OGG,
+            MIME_OPUS, MIME_PCM, MIME_WAV, MIME_WAVE, MIME_TGPP, MIME_PCM16,
+            MIME_OGG,
         ].includes?.(x.mime_type)) {
             return {
                 type: 'input_audio',
@@ -745,13 +693,14 @@ const packResp = async (resp, options) => {
     if (options?.raw) { return resp; }
     let [
         txt, audio, images, annotations, simpleText, annotationsMarkdown, end,
-        json, audioMimeType,
+        json, audioMimeType, audioSuffix,
     ] = [
             resp.text || '',                                                    // ChatGPT / Claude / Gemini / Ollama
             resp?.audio?.data,                                                  // ChatGPT audio mode
             resp?.images || [],                                                 // Gemini images via Openrouter
             resp?.references,                                                   // Gemini references
-            '', '', '', null, MIME_PCM16,
+            '', '', '', null, resp?.audio?.mimeType || resp?.audio?.mime || MIME_PCM16,
+            null,
         ];
     simpleText = txt;
     while ((end = getInfoEnd(simpleText))) {
@@ -762,10 +711,17 @@ const packResp = async (resp, options) => {
         const str = simpleText.indexOf(x);
         str >= 0 && (simpleText = simpleText.slice(0, str).trim());
     });
+    audioSuffix = audioMimeType === MIME_MPEGA
+        ? 'mp3' : audioMimeType.split('/').pop();
     audio = await ignoreErrFunc(async () => ({
-        data: await packPcmToWav(audio, {
-            input: BUFFER, expected: BUFFER, suffix: 'pcm.wav', ...options
-        }), mime: audioMimeType,
+        data: [MIME_PCM, MIME_PCM16].includes(audioMimeType)
+            ? await packPcmToWav(audio, {
+                input: BUFFER, expected: BUFFER, suffix: 'pcm.wav', ...options
+            })
+            : await convert(audio, {
+                input: BUFFER, expected: BUFFER, suffix: audioSuffix, ...options
+            }),
+        mime: audioMimeType, mimeType: audioMimeType,
     }));
     // bug fix: Google Gemini returns duplicate first image: {
     images && images.length > 1 && images.shift();
@@ -793,7 +749,6 @@ const packResp = async (resp, options) => {
     //             "title": "在線時鐘- 目前時間- 線上時鐘- 時鐘線上 - 鬧鐘",
     //             "url": "https://naozhong.tw/shijian/",
     //             "content": "- [鬧鐘](https://naozhong.tw/)\n- [計時器](https://naozhong.tw/jishiqi/)\n- [碼錶](https://naozhong.tw/miaobiao/)\n- [時間](https://naozhong.tw/shijian/)\n\n# 現在時間\n\n加入\n\n- [編輯](javascript:;)\n- [移至頂端](javascript:;)\n- [上移](javascript:;)\n- [下移](javascript:;)\n- [刪除](javascript:;)\n\n# 最常用\n\n| | |\n| --- | --- |\n| [台北](https://naozhong.tw/shijian/%E5%8F%B0%E5%8C%97/) | 10:09:14 |\n| [北京，中國](https://naozhong.tw/shijian/%E5%8C%97%E4%BA%AC-%E4%B8%AD%E5%9C%8B/) | 10:09:14 |\n| [上海，中國](https://naozhong.tw/shijian/%E4%B8%8A%E6%B5%B7-%E4%B8%AD%E5%9C%8B/) | 10:09:14 |\n| [烏魯木齊，中國](https://naozhong.tw/shijian/%E7%83%8F%E9%AD%AF%",
-    //             "dateTime": "2025-03-13 06:48:01" // jina deep-research only
     //         }
     //     },
     // ];
@@ -806,22 +761,6 @@ const packResp = async (resp, options) => {
         ).join('\n');
     }
     txt = txt.split('\n');
-    // const [reJinaStr, reJinaEnd]
-    //     = [`^\s*(${THINK_STR})`, `(${THINK_END})\s*$`].map(x => new RegExp(x));
-    // const fixJina = [];
-    // for (let l of txt) {
-    //     let catched = false;
-    //     if (reJinaStr.test(l)) {
-    //         fixJina.push(THINK_STR);
-    //         l = l.replace(reJinaStr, '');
-    //     }
-    //     if (reJinaEnd.test(l)) {
-    //         l = l.replace(reJinaEnd, '');
-    //         catched = true;
-    //     }
-    //     fixJina.push(l, ...catched ? [THINK_END, ''] : []);
-    // }
-    // txt = fixJina;
     // Fixed Markdown ``` not in new line for Gemini {
     for (let i in txt) {
         const reCode = new RegExp(`^(.+)(${MD_CODE})$`);
@@ -977,8 +916,8 @@ const mergeMsgs = (resp, calls) => [resp, ...calls.length ? [
     `⚠️ Tools recursion limit reached: ${MAX_TOOL_RECURSION}`
 ] : []].map(x => x.trim()).join('\n\n');
 
-const promptOpenRouter = async (aiId, content, options = {}) => {
-    let { provider, client, model } = await getAi(aiId);
+const promptOpenAI = async (aiId, content, options = {}) => {
+    let { provider, client, model, options: opts } = await getAi(aiId);
     let [
         result, resultAudio, resultImages, resultReasoning, event, resultTools,
         responded, modalities, source, reasoningEnd, reasoning_details,
@@ -998,11 +937,19 @@ const promptOpenRouter = async (aiId, content, options = {}) => {
     } else if (!modalities && model.image) {
         modalities = [TEXT, IMAGE];
     }
+    const audioConfig = options.audio || (
+        source === OPENAI && modalities?.find?.(x => x === AUDIO)
+        && { voice: DEFAULT_MODELS[OPENAI_VOICE], format: 'pcm16' }
+    );
+    const audioMimeType = model.name === LYRIA_3 ? MIME_MPEGA : MIME_PCM16;
     const googleImageMode = source === GOOGLE && modalities?.includes?.(IMAGE);
+    const imageConfig = model.name === GPT_IMAGE_2
+        ? { aspect_ratio: '16:9', image_size: '2K' } : googleImageMode
+            ? { aspect_ratio: '16:9', image_size: '4K' } : null;
     const packedTools = _tools.map(x => x.def);
     const ext = provider === OPENROUTER && !packedTools?.find(
         x => x.function.name === 'searchWeb'
-    ) && !options.jsonMode ? ONLINE : '';
+    ) && !options.jsonMode && !options.select ? ONLINE : ''; // `select` typically comes from a tool invocation.
     // if (provider === OPENAI) {
     //     // need more debug, currently openrouter is priority
     //     packedTools.push(...[
@@ -1022,20 +969,19 @@ const promptOpenRouter = async (aiId, content, options = {}) => {
         ]);
     }
     const resp = await client.chat.completions.create({
-        model: `${isOpenrouter(provider, model) ? `${source}/` : ''}${options.model}${ext}`,
-        ...history,
+        model: isOpenrouter(provider, model) ? (
+            `${source}/${options.model}${ext}`.toLowerCase()
+            + (opts?.preset || '')
+        ) : `${options.model}${ext}`, ...history,
         ...options.jsonMode ? { response_format: { type: JSON_OBJECT } } : {},
         ...provider === OLLAMA ? { keep_alive: -1 } : {},
-        modalities, audio: options.audio || (
-            modalities?.find?.(x => x === AUDIO)
-            && { voice: DEFAULT_MODELS[OPENAI_VOICE], format: 'pcm16' }
-        ), ...model?.tools ? {
+        modalities, ...audioConfig ? { audio: audioConfig } : {},
+        ...model?.tools ? {
             tools: options.tools ?? packedTools, tool_choice: 'auto',
-        } : {}, store: true, stream: true, ...googleImageMode ? {
-            // bug: image_size not working
+        } : {}, store: true, stream: true, ...imageConfig ? {
             // https://openrouter.ai/docs/features/multimodal/image-generation
             // https://openrouter.ai/google/gemini-3-pro-image-preview/api
-            image_config: { aspect_ratio: '16:9', image_size: '4K' },
+            image_config: imageConfig,
         } : {}, ...options.reasoning_effort ? {
             reasoning: { effort: options.reasoning_effort },
         } : {},
@@ -1104,14 +1050,18 @@ const promptOpenRouter = async (aiId, content, options = {}) => {
         (deltaText || deltaAudio?.length || deltaImages.length)
             && await streamResp({
                 text: options.delta ? deltaText : result,
-                ...respAudio.length ? { audio: { data: respAudio } } : {},
+                ...respAudio.length ? {
+                    audio: { data: respAudio, mimeType: audioMimeType }
+                } : {},
                 ...respImages.length ? { images: respImages } : {},
             }, options);
     }
     event = {
         ...assistant, text: result, tool_calls: resultTools,
         ...resultImages.length ? { images: resultImages } : {},
-        ...resultAudio.length ? { audio: { data: resultAudio } } : {},
+        ...resultAudio.length ? {
+            audio: { data: resultAudio, mimeType: audioMimeType }
+        } : {},
         ...annotations.length ? { annotations } : {},
     };
     switch (source) {
@@ -1129,7 +1079,7 @@ const promptOpenRouter = async (aiId, content, options = {}) => {
         = await handleToolsCall(event, { ...options, result });
     if (toolsResult.length
         && countToolCalls(toolsResponse) < MAX_TOOL_RECURSION) {
-        return promptOpenRouter(aiId, content, {
+        return promptOpenAI(aiId, content, {
             ...options, toolsResult, result: toolsResponse,
         });
     }
@@ -1138,11 +1088,11 @@ const promptOpenRouter = async (aiId, content, options = {}) => {
 };
 
 const promptGoogle = async (aiId, prompt, options = {}) => {
-    let { provider, client, model, options: opts } = await getAi(aiId);
+    let { provider, client, model } = await getAi(aiId);
     const target_model = options?.model || model.name;
     const M = MODELS[target_model];
     prompt = ensureString(prompt, { trim: true });
-    [IMAGEN_4_UPSCALE, LYRIA_002].includes(M.name) || assertPrompt(prompt);
+    [IMAGEN_4_UPSCALE, LYRIA_3].includes(M.name) || assertPrompt(prompt);
     M.tts && (prompt = `${options?.prompt || TTS_PROMPT}:\n\n${prompt}`);
     prompt = trimText(prompt, { limit: M.maxInputTokens });
     if (M.name === IMAGEN_4_UPSCALE) {
@@ -1236,30 +1186,38 @@ const promptGoogle = async (aiId, prompt, options = {}) => {
             };
         }
     } else if (M?.audio) { // https://ai.google.dev/gemini-api/docs/speech-generation#voices
-        if (M.name === LYRIA_002) {
-            const token = await opts.authClient.getAccessToken();
-            const region = 'us-central1'; // opts.location;
-            const url = `https://${region}-aiplatform.googleapis.com/v1/projects/${opts.project}/locations/${region}/publishers/google/models/${M.name}:predict`;
-            const payload = { instances: [{ prompt }] };
-            const response = await fetch(url, {
-                method: 'POST',
-                headers: {
-                    'Authorization': `Bearer ${token.token}`,
-                    'Content-Type': 'application/json'
-                },
-                body: JSON.stringify(payload)
+        if (M.name === LYRIA_3) {
+            var resp = await client.interactions.create({
+                model: M.name,
+                input: [{ type: TEXT, text: prompt }],
+                ...(options?.config ? { generation_config: options.config } : {}),
             });
-            if (!response.ok) {
-                throwError(`API Error: ${response.status} - ${await response.text()}`);
-            }
-            const data = await response.json();
-            var rawAudio = data?.predictions?.[0]?.bytesBase64Encoded;
-            assert(rawAudio, ERROR_GENERATING);
+            assert(!resp?.error, resp?.error?.message || ERROR_GENERATING);
+            if (options?.generateRaw) { return resp; }
+            await tryUntil(async () => {
+                if (resp?.status === 'completed') { return; }
+                assert(
+                    !['failed', 'cancelled', 'incomplete'].includes(resp?.status),
+                    `Google music generation failed: ${resp?.status}`,
+                );
+                resp = await client.interactions.get(resp.id);
+                assert(
+                    resp?.status === 'completed',
+                    `Waiting for Google music generation: ${resp.id}`,
+                );
+            }, { maxTry: 60 * 10, log });
+            const outputs = resp?.outputs || [];
+            const rawAudio = outputs.find(x => x?.type === AUDIO && x?.data);
+            const mimeType = rawAudio?.mime_type || rawAudio?.mimeType || MIME_MPEGA;
+            const suffix = mimeType === MIME_MPEGA ? 'mp3' : mimeType.split('/').pop();
+            assert(rawAudio?.data, ERROR_GENERATING);
             resp = {
-                text: '', audio: {
-                    data: await convert(rawAudio, {
-                        input: BASE64, suffix: 'wav', ...options || {},
-                    }), mimeType: MIME_WAV,
+                text: outputs.filter(x => x?.type === TEXT && x?.text)
+                    .map(x => x.text).join(LN1),
+                audio: {
+                    data: await convert(rawAudio.data, {
+                        input: BASE64, suffix, ...options || {},
+                    }), mimeType,
                 }, model: packModelId([provider, M.source, M.name]),
             };
         } else {
@@ -1352,97 +1310,6 @@ const promptGoogle = async (aiId, prompt, options = {}) => {
     return { ...resp, processing: false };
 };
 
-const promptOpenAI = async (aiId, prompt, options = {}) => {
-    let { provider, client, toFile, model } = await getAi(aiId);
-    const target_model = options?.model || model.name;
-    const M = MODELS[target_model];
-    prompt = ensureString(prompt, { trim: true });
-    if (M?.image) {
-        assertPrompt(prompt);
-        prompt = trimText(prompt, { limit: M.maxInputTokens });
-        var resp = await client.images.generate({
-            model: M.name, moderation: 'low', output_format: 'png',
-            prompt, quality: 'high', size: '1536x1024',
-            // n = 1, // 1 ~ 10
-            // @TODO: 支持 streaming image generation
-            // https://platform.openai.com/docs/guides/image-generation?api=responses
-            // Please note that right now, only gpt-image-1 and gpt-image-1-mini are
-            // supported in the Responses API. We're working on support for gpt-image-1.5.
-            // partial_images: 3, stream: true,
-            // @TODO: 支持 image editing {}
-            ...options?.params || {},
-        });
-        if (!options?.raw) {
-            // print(resp);
-            resp = {
-                text: '', images: await Promise.all(resp.data.map(
-                    async (x, i) => {
-                        const buffer = await convert(
-                            x.b64_json, { input: BASE64 }
-                        );
-                        return {
-                            data: await convert(buffer, {
-                                input: BUFFER, suffix: PNG_EXT, ...options || {}
-                            }),
-                            mimeType: (await getMime(
-                                buffer, `${i}.${resp.output_format}`
-                            ))?.mime || MIME_PNG,
-                        }
-                    })
-                ), model: packModelId([provider, M.source, M.name]),
-            };
-        }
-    } else if (M?.audio) {
-        assertPrompt(prompt);
-        const ins_prompt = options?.prompt || `${TTS_PROMPT}.`;
-        prompt = trimText(prompt, {
-            limit: M.maxInputTokens - countTokens(ins_prompt),
-        });
-        // https://platform.openai.com/docs/api-reference/audio/createSpeech
-        var resp = await client.audio.speech.create({
-            model: M.name, voice: DEFAULT_MODELS[OPENAI_VOICE],
-            instructions: ins_prompt, response_format: 'opus',
-            input: prompt, ...options?.params || {},
-        });
-        if (!options?.raw) {
-            resp = {
-                text: '', audio: {
-                    data: await convert(Buffer.from(
-                        await resp.arrayBuffer()
-                    ), { suffix: OGG_EXT, ...options || {} }),
-                    mimeType: MIME_OGG,
-                }, model: packModelId([provider, M.source, M.name]),
-            };
-        }
-    } else if (M?.hearing) {
-        const audio = options?.attachments?.[0]?.data || options?.attachments?.[0];
-        assert(audio, 'Audio attachment is required.');
-        const input = ensureString(options?.input, { case: 'UP' });
-        const { content, cleanup } = await convert(audio, {
-            input: options?.input, ...options || {}, expected: STREAM, INVALID_AUDIO,
-            suffix: ['', BUFFER].includes(input) ? OGG_EXT : null,
-            withCleanupFunc: true,
-        });
-        var resp = await client.audio.transcriptions.create({
-            file: await toFile(content), model: M.name,
-            response_format: 'text', ...options?.params || {},
-        });
-        await cleanup();
-        if (!options?.raw) {
-            resp = {
-                text: resp.trim(),
-                model: packModelId([provider, M.source, M.name]),
-            };
-        }
-    } else {
-        throwError('Unsupported model.');
-    }
-    await streamResp(
-        { ...resp, processing: true }, { ...options, noPack: true }
-    );
-    return { ...resp, processing: false };
-};
-
 const initChat = async (options = {}) => {
     if (options.sessions) {
         assert(
@@ -1461,7 +1328,7 @@ const initChat = async (options = {}) => {
         countTokens(chatConfig.systemPrompt), await getAi(null, { all: true })
     ];
     for (const ai of ais.filter(x => ![
-        IMAGEN_4_UPSCALE, VEO_31, GPT_4O_TRANSCRIBE, LYRIA_002,
+        IMAGEN_4_UPSCALE, VEO_31,
     ].includes(x.model.name))) {
         const mxPmpt = ai.model.maxInputTokens / 2;
         assert(spTokens < mxPmpt,
@@ -1666,15 +1533,17 @@ const distillFile = async (attachments, o) => {
 };
 
 const tts = async (content, options = {}) => {
-    const resp = await prompt(
-        content, { select: { audio: true, fast: true }, ...options, raw: false }
-    );
+    const resp = await prompt(content, {
+        select: { audio: true, fast: true }, ...options, raw: false,
+        audioMode: true,
+    });
     return options.raw ? resp : resp?.audio?.data;
 };
 
-const stt = async (audio, options = {}) => await distillFile(
-    audio, { prompt: STT_PROMPT, ...options }
-);
+const stt = async (audio, options = {}) => await distillFile(audio, {
+    select: { hearing: true, fast: true },
+    prompt: STT_PROMPT, ...options
+});
 
 const prompt = async (input, options = {}) => {
     const ai = await getAi(options?.aiId, options);
@@ -1789,19 +1658,18 @@ export default init;
 export {
     _NEED,
     _NO_RENDER,
-    CLOUD_OPUS_46,
+    CLOUD_OPUS_47,
     CODE_INTERPRETER,
-    DEEPSEEK_32,
     DEFAULT_MODELS,
     FEATURE_ICONS,
     FUNCTION,
-    GEMINI_25_FLASH_TTS,
     GEMINI_30_FLASH,
     GEMINI_25_PRO_TTS,
     GEMINI_30_PRO_IMAGE,
     GEMINI_31_PRO,
-    GPT_IMAGE_15,
-    GPT_54,
+    GPT_IMAGE_2,
+    GPT_55,
+    GPT_55_MINI,
     INSTRUCTIONS,
     MODELS,
     OPENAI_VOICE,
@@ -1822,7 +1690,7 @@ export {
     k,
     listOpenAIModels,
     prompt,
-    promptOpenRouter,
+    promptOpenAI,
     resetSession,
     setSession,
     stt,