@hiper2d/ai-agents 0.15.0 → 0.16.0

package/dist/index.mjs

@@ -1756,8 +1756,9 @@ function extractTokenUsageFromResponse8(response) {
 var VOICE_MODEL_CONSTANTS = {
   OPENAI_TTS: "gpt-4o-mini-tts",
   OPENAI_STT: "whisper-1",
-  // ai.google.dev/gemini-api/docs/speech-generation
-  GOOGLE_TTS: "gemini-3.1-flash-tts-preview",
+  // ai.google.dev/gemini-api/docs/speech-generation — Google's named replacement
+  // for gemini-3.1-flash-tts-preview (2026-09-24: ~6 s vs ~15 s per line, ~4x cheaper)
+  GOOGLE_TTS: "gemini-3.8-flash-lite-tts",
   // ai.google.dev/gemini-api/docs/transcribe — Interactions API only (see google-stt.ts)
   GOOGLE_STT: "gemini-3.5-transcribe"
 };
@@ -1769,9 +1770,9 @@ var VOICE_PROVIDER_API_KEY = {
 var VOICE_MODEL_PRICING = {
   [VOICE_MODEL_CONSTANTS.OPENAI_TTS]: { pricePerMillionCharacters: 15 },
   [VOICE_MODEL_CONSTANTS.OPENAI_STT]: { pricePerMinute: 6e-3 },
-  // Measured 2026-09-05: ~32 audio tokens per second of speech, so a 15-second
-  // line is ~$0.01 — about 3-5x an OpenAI line of the same length.
-  [VOICE_MODEL_CONSTANTS.GOOGLE_TTS]: { textInputPricePerM: 1, audioOutputPricePerM: 20 },
+  // ~32 audio tokens per second of speech, so a 15-second line is ~$0.003.
+  // Google doubles both rates to $1 / $12 on 2027-01-01 — update then.
+  [VOICE_MODEL_CONSTANTS.GOOGLE_TTS]: { textInputPricePerM: 0.5, audioOutputPricePerM: 6 },
   // ~25 audio tokens per second in, ~175 text tokens per minute out: ≈ $0.005/min.
   [VOICE_MODEL_CONSTANTS.GOOGLE_STT]: { audioInputPricePerM: 2, textOutputPricePerM: 12 }
 };
@@ -1878,12 +1879,10 @@ import { GoogleGenAI } from "@google/genai";
 var AUDIO_TOKENS_PER_SECOND = 32;
 var SAMPLE_RATE = 24e3;
 var PCM_BYTES_PER_SECOND = SAMPLE_RATE * 2;
-function buildGoogleTtsPrompt(text, voiceStyle) {
+var WAV_HEADER_BYTES = 44;
+function normalizeTtsStyle(voiceStyle) {
   const style = voiceStyle?.trim().replace(/[:.!,;\s]+$/, "");
-  if (!style) return text;
-  const isShort = style.split(/\s+/).length <= 3 && !/[.!?,;]/.test(style);
-  return isShort ? `Say ${style}: ${text}` : `${style}:
-${text}`;
+  return style || void 0;
 }
 function pcmToWav(pcmData, sampleRate = SAMPLE_RATE) {
   const numChannels = 1;
@@ -1927,28 +1926,51 @@ function describeEmptyTtsResponse(response) {
   if (!parts.length) bits.push("no parts");
   return bits.join(", ");
 }
+function isSafetyBlocked(response) {
+  const finish = response?.candidates?.[0]?.finishReason;
+  return finish === "SAFETY" || finish === "PROHIBITED_CONTENT" || !!response?.promptFeedback?.blockReason;
+}
+function findAudioPart(response) {
+  const parts = response?.candidates?.[0]?.content?.parts ?? [];
+  const inline = parts.find((part) => part.inlineData?.mimeType?.startsWith("audio/") && part.inlineData?.data)?.inlineData;
+  return inline ? { data: inline.data, mimeType: inline.mimeType } : void 0;
+}
 async function generateGoogleTtsAudio(text, apiKey, options) {
   const client = new GoogleGenAI({ apiKey });
-  const response = await client.models.generateContent({
+  const request = (style2) => client.models.generateContent({
     model: VOICE_MODEL_CONSTANTS.GOOGLE_TTS,
-    contents: [{ parts: [{ text: buildGoogleTtsPrompt(text, options.voiceStyle) }] }],
+    contents: [{ parts: [style2 ? { text, speechMetadata: { style: style2 } } : { text }] }],
     config: {
       responseModalities: ["AUDIO"],
       speechConfig: { voiceConfig: { prebuiltVoiceConfig: { voiceName: options.voiceName } } }
     }
   });
-  const candidate = response.candidates?.[0];
-  const parts = candidate?.content?.parts ?? [];
-  const audioPart = parts.find((part) => part.inlineData?.mimeType?.startsWith("audio/"));
-  if (!audioPart?.inlineData?.data) {
-    throw new Error(`No audio data in Google TTS response (${describeEmptyTtsResponse(response)})`);
-  }
-  const pcmData = new Uint8Array(Buffer.from(audioPart.inlineData.data, "base64"));
+  const style = normalizeTtsStyle(options.voiceStyle);
+  let response = await request(style);
+  let styleDropped = false;
+  let blockedInputTokens = 0;
+  if (!findAudioPart(response) && style && isSafetyBlocked(response)) {
+    blockedInputTokens = response?.usageMetadata?.promptTokenCount ?? 0;
+    response = await request(void 0);
+    styleDropped = true;
+  }
+  const audioPart = findAudioPart(response);
+  if (!audioPart) {
+    throw new Error(`No audio data in Google TTS response (${describeEmptyTtsResponse(response)}${styleDropped ? ", also without the style" : ""})`);
+  }
+  const bytes = new Uint8Array(Buffer.from(audioPart.data, "base64"));
+  const isWav = /^audio\/(x-)?wav/i.test(audioPart.mimeType);
+  const audio = isWav ? bytes.buffer.slice(bytes.byteOffset, bytes.byteOffset + bytes.byteLength) : pcmToWav(bytes);
+  const pcmBytes = isWav ? Math.max(0, bytes.length - WAV_HEADER_BYTES) : bytes.length;
   const usageMetadata = response.usageMetadata ?? {};
-  const inputTokens = usageMetadata.promptTokenCount ?? 0;
+  const inputTokens = (usageMetadata.promptTokenCount ?? 0) + blockedInputTokens;
   const reportedOutput = usageMetadata.candidatesTokenCount;
-  const outputTokens = reportedOutput && reportedOutput > 0 ? reportedOutput : Math.ceil(pcmData.length / PCM_BYTES_PER_SECOND * AUDIO_TOKENS_PER_SECOND);
-  return { audio: pcmToWav(pcmData), usage: { inputTokens, outputTokens } };
+  const outputTokens = reportedOutput && reportedOutput > 0 ? reportedOutput : Math.ceil(pcmBytes / PCM_BYTES_PER_SECOND * AUDIO_TOKENS_PER_SECOND);
+  return {
+    audio,
+    usage: { inputTokens, outputTokens },
+    ...styleDropped ? { styleDropped } : {}
+  };
 }
 
 // src/voice/google-stt.ts
@@ -1979,11 +2001,11 @@ var GoogleVoiceAgent = class {
   ttsModel = VOICE_MODEL_CONSTANTS.GOOGLE_TTS;
   sttModel = VOICE_MODEL_CONSTANTS.GOOGLE_STT;
   async speak(request) {
-    const { audio, usage } = await generateGoogleTtsAudio(request.text, this.apiKey, {
+    const { audio, usage, styleDropped } = await generateGoogleTtsAudio(request.text, this.apiKey, {
       voiceName: request.voice,
       voiceStyle: request.voiceStyle
     });
-    return { audio, costUSD: calculateGeminiTtsCost(usage), usage };
+    return { audio, costUSD: calculateGeminiTtsCost(usage), usage, ...styleDropped ? { styleDropped } : {} };
   }
   async transcribe(request) {
     const { text, durationSeconds, usage } = await transcribeWithGemini(request.audio, this.apiKey, { mimeType: request.mimeType });
@@ -5351,7 +5373,6 @@ export {
   XAI_REASONING_EFFORTS,
   ZodSchemaConverter,
   applySpend,
-  buildGoogleTtsPrompt,
   calculateAnthropicCost,
   calculateCost,
   calculateDeepSeekCost,
@@ -5404,6 +5425,7 @@ export {
   isHybridThinkingModel,
   isInPeakWindow,
   isPeakBilling,
+  isSafetyBlocked,
   isWeekendAt,
   ledgerSpend,
   logger,
@@ -5411,6 +5433,7 @@ export {
   modelHasTag,
   modelIsFast,
   needsPromptBasedSchema,
+  normalizeTtsStyle,
   parseAndValidateLlmJson,
   pcmToWav,
   periodEnd,