getpatter 0.4.4 → 0.5.1

package/dist/{chunk-O3RQG3NL.mjs → chunk-B6C3KIBG.mjs}

@@ -247,242 +247,13 @@ var ElevenLabsConvAIAdapter = class {
   }
 };
 
-// src/providers/deepgram-stt.ts
-import WebSocket3 from "ws";
-var DEEPGRAM_WS_URL = "wss://api.deepgram.com/v1/listen";
-var DeepgramSTT = class _DeepgramSTT {
-  ws = null;
-  callbacks = [];
-  /** Request ID from Deepgram — used to query actual cost post-call. */
-  requestId = "";
-  apiKey;
-  language;
-  model;
-  encoding;
-  sampleRate;
-  endpointingMs;
-  utteranceEndMs;
-  smartFormat;
-  interimResults;
-  vadEvents;
-  constructor(apiKey, languageOrOptions, model, encoding, sampleRate, options) {
-    this.apiKey = apiKey;
-    const opts = typeof languageOrOptions === "object" && languageOrOptions !== null ? languageOrOptions : options ?? {};
-    this.language = (typeof languageOrOptions === "string" ? languageOrOptions : opts.language) ?? "en";
-    this.model = model ?? opts.model ?? "nova-3";
-    this.encoding = encoding ?? opts.encoding ?? "linear16";
-    this.sampleRate = sampleRate ?? opts.sampleRate ?? 16e3;
-    this.endpointingMs = opts.endpointingMs ?? 150;
-    this.utteranceEndMs = opts.utteranceEndMs === null ? null : opts.utteranceEndMs ?? 1e3;
-    this.smartFormat = opts.smartFormat ?? true;
-    this.interimResults = opts.interimResults ?? true;
-    this.vadEvents = opts.vadEvents ?? true;
-  }
-  /** Factory for Twilio calls — mulaw 8 kHz. Forwards tuning options through. */
-  static forTwilio(apiKey, language = "en", model = "nova-3", options = {}) {
-    return new _DeepgramSTT(apiKey, language, model, "mulaw", 8e3, options);
-  }
-  async connect() {
-    const params = new URLSearchParams({
-      model: this.model,
-      language: this.language,
-      encoding: this.encoding,
-      sample_rate: String(this.sampleRate),
-      channels: "1",
-      interim_results: this.interimResults ? "true" : "false",
-      endpointing: String(this.endpointingMs),
-      smart_format: this.smartFormat ? "true" : "false",
-      vad_events: this.vadEvents ? "true" : "false",
-      no_delay: "true"
-    });
-    if (this.utteranceEndMs !== null) {
-      params.set("utterance_end_ms", String(Math.max(this.utteranceEndMs, 1e3)));
-    }
-    const url = `${DEEPGRAM_WS_URL}?${params.toString()}`;
-    this.ws = new WebSocket3(url, {
-      headers: { Authorization: `Token ${this.apiKey}` }
-    });
-    await new Promise((resolve, reject) => {
-      const timer = setTimeout(() => reject(new Error("Deepgram connect timeout")), 1e4);
-      this.ws.once("open", () => {
-        clearTimeout(timer);
-        resolve();
-      });
-      this.ws.once("error", (err) => {
-        clearTimeout(timer);
-        reject(err);
-      });
-    });
-    this.ws.on("message", (raw) => {
-      let data;
-      try {
-        data = JSON.parse(raw.toString());
-      } catch {
-        return;
-      }
-      if (data.type === "Metadata" && data.request_id) {
-        this.requestId = data.request_id;
-        return;
-      }
-      if (data.type !== "Results") return;
-      const alternatives = data.channel?.alternatives ?? [];
-      if (!alternatives.length) return;
-      const best = alternatives[0];
-      const text = (best.transcript ?? "").trim();
-      if (!text) return;
-      const transcript = {
-        text,
-        isFinal: Boolean(data.is_final) || Boolean(data.speech_final),
-        confidence: best.confidence ?? 0
-      };
-      for (const cb of this.callbacks) {
-        cb(transcript);
-      }
-    });
-  }
-  sendAudio(audio) {
-    if (!this.ws || this.ws.readyState !== WebSocket3.OPEN) return;
-    this.ws.send(audio);
-  }
-  onTranscript(callback) {
-    if (this.callbacks.length >= 10) {
-      getLogger().warn("DeepgramSTT: maximum of 10 onTranscript callbacks reached; replacing the last callback.");
-      this.callbacks[this.callbacks.length - 1] = callback;
-      return;
-    }
-    this.callbacks.push(callback);
-  }
-  close() {
-    if (this.ws) {
-      try {
-        this.ws.send(JSON.stringify({ type: "CloseStream" }));
-      } catch {
-      }
-      this.ws.close();
-      this.ws = null;
-    }
-  }
-};
-
-// src/providers/whisper-stt.ts
-var OPENAI_TRANSCRIPTION_URL = "https://api.openai.com/v1/audio/transcriptions";
-var DEFAULT_BUFFER_SIZE = 16e3 * 2;
-function wrapPcmInWav(pcm, sampleRate = 16e3, channels = 1, bitsPerSample = 16) {
-  const dataSize = pcm.length;
-  const header = Buffer.alloc(44);
-  header.write("RIFF", 0);
-  header.writeUInt32LE(36 + dataSize, 4);
-  header.write("WAVE", 8);
-  header.write("fmt ", 12);
-  header.writeUInt32LE(16, 16);
-  header.writeUInt16LE(1, 20);
-  header.writeUInt16LE(channels, 22);
-  header.writeUInt32LE(sampleRate, 24);
-  header.writeUInt32LE(sampleRate * channels * (bitsPerSample / 8), 28);
-  header.writeUInt16LE(channels * (bitsPerSample / 8), 32);
-  header.writeUInt16LE(bitsPerSample, 34);
-  header.write("data", 36);
-  header.writeUInt32LE(dataSize, 40);
-  return Buffer.concat([header, pcm]);
+// src/provider-factory.ts
+async function createSTT(agent) {
+  return agent.stt ?? null;
+}
+async function createTTS(agent) {
+  return agent.tts ?? null;
 }
-var WhisperSTT = class _WhisperSTT {
-  apiKey;
-  model;
-  language;
-  bufferSize;
-  buffer = Buffer.alloc(0);
-  callbacks = [];
-  running = false;
-  pendingTranscriptions = [];
-  constructor(apiKey, model = "whisper-1", language, bufferSize = DEFAULT_BUFFER_SIZE) {
-    this.apiKey = apiKey;
-    this.model = model;
-    this.language = language;
-    this.bufferSize = bufferSize;
-  }
-  /** Factory for Twilio calls — mulaw 8 kHz is transcoded upstream, so we still receive PCM 16-bit. */
-  static forTwilio(apiKey, language = "en", model = "whisper-1") {
-    return new _WhisperSTT(apiKey, model, language);
-  }
-  async connect() {
-    this.running = true;
-    this.buffer = Buffer.alloc(0);
-  }
-  sendAudio(audio) {
-    if (!this.running) return;
-    this.buffer = Buffer.concat([this.buffer, audio]);
-    if (this.buffer.length >= this.bufferSize) {
-      const pcm = this.buffer;
-      this.buffer = Buffer.alloc(0);
-      this.trackTranscription(this.transcribeBuffer(pcm));
-    }
-  }
-  trackTranscription(promise) {
-    const wrapped = promise.finally(() => {
-      const idx = this.pendingTranscriptions.indexOf(wrapped);
-      if (idx !== -1) this.pendingTranscriptions.splice(idx, 1);
-    });
-    this.pendingTranscriptions.push(wrapped);
-  }
-  onTranscript(callback) {
-    if (this.callbacks.length >= 10) {
-      getLogger().warn("WhisperSTT: maximum of 10 onTranscript callbacks reached; replacing the last callback.");
-      this.callbacks[this.callbacks.length - 1] = callback;
-      return;
-    }
-    this.callbacks.push(callback);
-  }
-  async close() {
-    this.running = false;
-    if (this.buffer.length >= this.bufferSize / 4) {
-      const pcm = this.buffer;
-      this.buffer = Buffer.alloc(0);
-      this.trackTranscription(this.transcribeBuffer(pcm));
-    } else {
-      this.buffer = Buffer.alloc(0);
-    }
-    await Promise.allSettled(this.pendingTranscriptions);
-    this.callbacks = [];
-  }
-  // ------------------------------------------------------------------
-  // Private
-  // ------------------------------------------------------------------
-  async transcribeBuffer(pcm) {
-    const wav = wrapPcmInWav(pcm);
-    const formData = new FormData();
-    formData.append("file", new Blob([wav.buffer.slice(wav.byteOffset, wav.byteOffset + wav.byteLength)], { type: "audio/wav" }), "audio.wav");
-    formData.append("model", this.model);
-    if (this.language) {
-      formData.append("language", this.language);
-    }
-    try {
-      const resp = await fetch(OPENAI_TRANSCRIPTION_URL, {
-        method: "POST",
-        headers: { Authorization: `Bearer ${this.apiKey}` },
-        body: formData,
-        signal: AbortSignal.timeout(15e3)
-      });
-      if (!resp.ok) {
-        const body = await resp.text();
-        getLogger().error(`WhisperSTT transcription error: ${resp.status} ${body}`);
-        return;
-      }
-      const json = await resp.json();
-      const text = (json.text ?? "").trim();
-      if (!text) return;
-      const transcript = {
-        text,
-        isFinal: true,
-        confidence: 1
-      };
-      for (const cb of this.callbacks) {
-        cb(transcript);
-      }
-    } catch (err) {
-      getLogger().error(`WhisperSTT transcription error: ${String(err)}`);
-    }
-  }
-};
 
 // src/pricing.ts
 var DEFAULT_PRICING = {
@@ -1855,248 +1626,120 @@ function isWebSocketUrl(url) {
   return url.startsWith("ws://") || url.startsWith("wss://");
 }
 
-// src/providers/elevenlabs-tts.ts
-var ELEVENLABS_BASE_URL = "https://api.elevenlabs.io/v1";
-var ELEVENLABS_VOICE_ID_BY_NAME = {
-  rachel: "21m00Tcm4TlvDq8ikWAM",
-  drew: "29vD33N1CtxCmqQRPOHJ",
-  clyde: "2EiwWnXFnvU5JabPnv8n",
-  paul: "5Q0t7uMcjvnagumLfvZi",
-  domi: "AZnzlk1XvdvUeBnXmlld",
-  dave: "CYw3kZ02Hs0563khs1Fj",
-  fin: "D38z5RcWu1voky8WS1ja",
-  bella: "EXAVITQu4vr4xnSDxMaL",
-  antoni: "ErXwobaYiN019PkySvjV",
-  thomas: "GBv7mTt0atIp3Br8iCZE",
-  charlie: "IKne3meq5aSn9XLyUdCD",
-  george: "JBFqnCBsd6RMkjVDRZzb",
-  emily: "LcfcDJNUP1GQjkzn1xUU",
-  elli: "MF3mGyEYCl7XYWbV9V6O",
-  callum: "N2lVS1w4EtoT3dr4eOWO",
-  patrick: "ODq5zmih8GrVes37Dizd",
-  harry: "SOYHLrjzK2X1ezoPC6cr",
-  liam: "TX3LPaxmHKxFdv7VOQHJ",
-  dorothy: "ThT5KcBeYPX3keUQqHPh",
-  josh: "TxGEqnHWrfWFTfGW9XjX",
-  arnold: "VR6AewLTigWG4xSOukaG",
-  charlotte: "XB0fDUnXU5powFXDhCwa",
-  matilda: "XrExE9yKIg1WjnnlVkGX",
-  matthew: "Yko7PKHZNXotIFUBG7I9",
-  james: "ZQe5CZNOzWyzPSCn5a3c",
-  joseph: "Zlb1dXrM653N07WRdFW3",
-  jeremy: "bVMeCyTHy58xNoL34h3p",
-  michael: "flq6f7yk4E4fJM5XTYuZ",
-  ethan: "g5CIjZEefAph4nQFvHAz",
-  gigi: "jBpfuIE2acCO8z3wKNLl",
-  freya: "jsCqWAovK2LkecY7zXl4",
-  brian: "nPczCjzI2devNBz1zQrb",
-  grace: "oWAxZDx7w5VEj9dCyTzz",
-  daniel: "onwK4e9ZLuTAKqWW03F9",
-  lily: "pFZP5JQG7iQjIQuC4Bku",
-  serena: "pMsXgVXv3BLzUgSXRplE",
-  adam: "pNInz6obpgDQGcFmaJgB",
-  nicole: "piTKgcLEGmPE4e6mEKli",
-  bill: "pqHfZKP75CvOlQylNhV4",
-  jessie: "t0jbNlBVZ17f02VDIeMI",
-  ryan: "wViXBPUzp2ZZixB1xQuM",
-  sam: "yoZ06aMxZJJ28mfd3POQ",
-  glinda: "z9fAnlkpzviPz146aGWa",
-  giovanni: "zcAOhNBS3c14rBihAFp1",
-  mimi: "zrHiDhphv9ZnVXBqCLjz",
-  alloy: "21m00Tcm4TlvDq8ikWAM"
-};
-var VOICE_ID_PATTERN = /^[A-Za-z0-9]{20}$/;
-function resolveVoiceId(voice) {
-  if (!voice) return voice;
-  if (VOICE_ID_PATTERN.test(voice)) return voice;
-  return ELEVENLABS_VOICE_ID_BY_NAME[voice.toLowerCase()] ?? voice;
-}
-var ElevenLabsTTS = class {
-  constructor(apiKey, voiceId = "21m00Tcm4TlvDq8ikWAM", modelId = "eleven_turbo_v2_5", outputFormat = "pcm_16000") {
+// src/providers/deepgram-stt.ts
+import WebSocket3 from "ws";
+var DEEPGRAM_WS_URL = "wss://api.deepgram.com/v1/listen";
+var DeepgramSTT = class _DeepgramSTT {
+  ws = null;
+  callbacks = [];
+  /** Request ID from Deepgram — used to query actual cost post-call. */
+  requestId = "";
+  apiKey;
+  language;
+  model;
+  encoding;
+  sampleRate;
+  endpointingMs;
+  utteranceEndMs;
+  smartFormat;
+  interimResults;
+  vadEvents;
+  constructor(apiKey, languageOrOptions, model, encoding, sampleRate, options) {
     this.apiKey = apiKey;
-    this.modelId = modelId;
-    this.outputFormat = outputFormat;
-    this.voiceId = resolveVoiceId(voiceId);
+    const opts = typeof languageOrOptions === "object" && languageOrOptions !== null ? languageOrOptions : options ?? {};
+    this.language = (typeof languageOrOptions === "string" ? languageOrOptions : opts.language) ?? "en";
+    this.model = model ?? opts.model ?? "nova-3";
+    this.encoding = encoding ?? opts.encoding ?? "linear16";
+    this.sampleRate = sampleRate ?? opts.sampleRate ?? 16e3;
+    this.endpointingMs = opts.endpointingMs ?? 150;
+    this.utteranceEndMs = opts.utteranceEndMs === null ? null : opts.utteranceEndMs ?? 1e3;
+    this.smartFormat = opts.smartFormat ?? true;
+    this.interimResults = opts.interimResults ?? true;
+    this.vadEvents = opts.vadEvents ?? true;
   }
-  voiceId;
-  /**
-   * Synthesise text to speech and return the full audio as a single Buffer.
-   *
-   * For large chunks (or when latency matters) call `synthesizeStream` instead.
-   */
-  async synthesize(text) {
-    const chunks = [];
-    for await (const chunk of this.synthesizeStream(text)) {
-      chunks.push(chunk);
-    }
-    return Buffer.concat(chunks);
+  /** Factory for Twilio calls — mulaw 8 kHz. Forwards tuning options through. */
+  static forTwilio(apiKey, language = "en", model = "nova-3", options = {}) {
+    return new _DeepgramSTT(apiKey, language, model, "mulaw", 8e3, options);
   }
-  /**
-   * Synthesise text and yield audio chunks as they arrive (streaming).
-   *
-   * The yielded buffers are raw PCM at 16 kHz (or whatever `outputFormat` is
-   * configured to).
-   */
-  async *synthesizeStream(text) {
-    const url = `${ELEVENLABS_BASE_URL}/text-to-speech/${encodeURIComponent(this.voiceId)}/stream?output_format=${encodeURIComponent(this.outputFormat)}`;
-    const response = await fetch(url, {
-      method: "POST",
-      headers: {
-        "xi-api-key": this.apiKey,
-        "Content-Type": "application/json"
-      },
-      body: JSON.stringify({ text, model_id: this.modelId }),
-      signal: AbortSignal.timeout(3e4)
+  async connect() {
+    const params = new URLSearchParams({
+      model: this.model,
+      language: this.language,
+      encoding: this.encoding,
+      sample_rate: String(this.sampleRate),
+      channels: "1",
+      interim_results: this.interimResults ? "true" : "false",
+      endpointing: String(this.endpointingMs),
+      smart_format: this.smartFormat ? "true" : "false",
+      vad_events: this.vadEvents ? "true" : "false",
+      no_delay: "true"
     });
-    if (!response.ok) {
-      const body = await response.text();
-      throw new Error(`ElevenLabs TTS error ${response.status}: ${body}`);
-    }
-    if (!response.body) {
-      throw new Error("ElevenLabs TTS: no response body");
+    if (this.utteranceEndMs !== null) {
+      params.set("utterance_end_ms", String(Math.max(this.utteranceEndMs, 1e3)));
     }
-    const reader = response.body.getReader();
-    try {
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value && value.length > 0) {
-          yield Buffer.from(value);
-        }
-      }
-    } finally {
-      if (typeof reader.cancel === "function") await reader.cancel().catch(() => {
+    const url = `${DEEPGRAM_WS_URL}?${params.toString()}`;
+    this.ws = new WebSocket3(url, {
+      headers: { Authorization: `Token ${this.apiKey}` }
+    });
+    await new Promise((resolve, reject) => {
+      const timer = setTimeout(() => reject(new Error("Deepgram connect timeout")), 1e4);
+      this.ws.once("open", () => {
+        clearTimeout(timer);
+        resolve();
+      });
+      this.ws.once("error", (err) => {
+        clearTimeout(timer);
+        reject(err);
       });
-      reader.releaseLock();
-    }
-  }
-};
-
-// src/providers/openai-tts.ts
-var OPENAI_TTS_URL = "https://api.openai.com/v1/audio/speech";
-var OpenAITTS = class _OpenAITTS {
-  constructor(apiKey, voice = "alloy", model = "tts-1") {
-    this.apiKey = apiKey;
-    this.voice = voice;
-    this.model = model;
-  }
-  /**
-   * Synthesise text to speech and return the full audio as a single Buffer.
-   *
-   * For large chunks (or when latency matters) call `synthesizeStream` instead.
-   */
-  async synthesize(text) {
-    const chunks = [];
-    for await (const chunk of this.synthesizeStream(text)) {
-      chunks.push(chunk);
-    }
-    return Buffer.concat(chunks);
-  }
-  /**
-   * Synthesise text and yield audio chunks as they arrive (streaming).
-   *
-   * OpenAI returns 24 kHz PCM16; each chunk is resampled to 16 kHz before
-   * yielding so the output is ready for telephony pipelines.
-   *
-   * The resampler carries state (buffered samples + odd trailing byte)
-   * between chunks — without that state cross-chunk sample alignment drifts
-   * and the caller hears pops / dropped audio (BUG #23, mirror of the
-   * Python `audioop.ratecv` fix).
-   */
-  async *synthesizeStream(text) {
-    const response = await fetch(OPENAI_TTS_URL, {
-      method: "POST",
-      headers: {
-        "Authorization": `Bearer ${this.apiKey}`,
-        "Content-Type": "application/json"
-      },
-      body: JSON.stringify({
-        model: this.model,
-        input: text,
-        voice: this.voice,
-        response_format: "pcm"
-      }),
-      signal: AbortSignal.timeout(3e4)
     });
-    if (!response.ok) {
-      const body = await response.text();
-      throw new Error(`OpenAI TTS error ${response.status}: ${body}`);
-    }
-    if (!response.body) {
-      throw new Error("OpenAI TTS: no response body");
-    }
-    const ctx = { carryByte: null, leftover: [] };
-    const reader = response.body.getReader();
-    try {
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        if (value && value.length > 0) {
-          const out = _OpenAITTS.resampleStreaming(Buffer.from(value), ctx);
-          if (out.length > 0) yield out;
-        }
+    this.ws.on("message", (raw) => {
+      let data;
+      try {
+        data = JSON.parse(raw.toString());
+      } catch {
+        return;
       }
-      if (ctx.leftover.length > 0) {
-        const tail = Buffer.alloc(ctx.leftover.length * 2);
-        for (let i = 0; i < ctx.leftover.length; i++) {
-          tail.writeInt16LE(ctx.leftover[i], i * 2);
-        }
-        yield tail;
+      if (data.type === "Metadata" && data.request_id) {
+        this.requestId = data.request_id;
+        return;
       }
-    } finally {
-      if (typeof reader.cancel === "function") await reader.cancel().catch(() => {
-      });
-      reader.releaseLock();
-    }
+      if (data.type !== "Results") return;
+      const alternatives = data.channel?.alternatives ?? [];
+      if (!alternatives.length) return;
+      const best = alternatives[0];
+      const text = (best.transcript ?? "").trim();
+      if (!text) return;
+      const transcript = {
+        text,
+        isFinal: Boolean(data.is_final) || Boolean(data.speech_final),
+        confidence: best.confidence ?? 0
+      };
+      for (const cb of this.callbacks) {
+        cb(transcript);
+      }
+    });
   }
-  /**
-   * Streaming 24 kHz → 16 kHz resampler (PCM16-LE). Maintains cross-chunk
-   * state so the 3:2 pattern doesn't reset at every network read.
-   */
-  static resampleStreaming(audio, ctx) {
-    let buf;
-    if (ctx.carryByte !== null) {
-      buf = Buffer.concat([Buffer.from([ctx.carryByte]), audio]);
-      ctx.carryByte = null;
-    } else {
-      buf = audio;
-    }
-    if (buf.length % 2 === 1) {
-      ctx.carryByte = buf[buf.length - 1];
-      buf = buf.subarray(0, buf.length - 1);
-    }
-    if (buf.length === 0 && ctx.leftover.length === 0) {
-      return Buffer.alloc(0);
-    }
-    const sampleCount = buf.length / 2;
-    const samples = ctx.leftover.slice();
-    for (let i2 = 0; i2 < sampleCount; i2++) {
-      samples.push(buf.readInt16LE(i2 * 2));
-    }
-    const out = [];
-    let i = 0;
-    while (i + 2 < samples.length) {
-      out.push(samples[i]);
-      out.push(Math.trunc((samples[i + 1] + samples[i + 2]) / 2));
-      i += 3;
-    }
-    ctx.leftover = samples.slice(i);
-    const buffer = Buffer.alloc(out.length * 2);
-    for (let j = 0; j < out.length; j++) {
-      buffer.writeInt16LE(out[j], j * 2);
+  sendAudio(audio) {
+    if (!this.ws || this.ws.readyState !== WebSocket3.OPEN) return;
+    this.ws.send(audio);
+  }
+  onTranscript(callback) {
+    if (this.callbacks.length >= 10) {
+      getLogger().warn("DeepgramSTT: maximum of 10 onTranscript callbacks reached; replacing the last callback.");
+      this.callbacks[this.callbacks.length - 1] = callback;
+      return;
     }
-    return buffer;
+    this.callbacks.push(callback);
   }
-  /** @deprecated use {@link resampleStreaming} with persistent state. */
-  static resample24kTo16k(audio) {
-    const ctx = { carryByte: null, leftover: [] };
-    const out = _OpenAITTS.resampleStreaming(audio, ctx);
-    if (ctx.leftover.length === 0) return out;
-    const tail = Buffer.alloc(ctx.leftover.length * 2);
-    for (let i = 0; i < ctx.leftover.length; i++) {
-      tail.writeInt16LE(ctx.leftover[i], i * 2);
+  close() {
+    if (this.ws) {
+      try {
+        this.ws.send(JSON.stringify({ type: "CloseStream" }));
+      } catch {
+      }
+      this.ws.close();
+      this.ws = null;
     }
-    return Buffer.concat([out, tail]);
   }
 };
 
@@ -2697,8 +2340,8 @@ var StreamHandler = class {
     this.caller = caller;
     this.callee = callee;
     this.history = createHistoryManager(200);
-    const sttProviderName = deps.agent.stt?.provider || (deps.agent.deepgramKey ? "deepgram" : void 0);
-    const ttsProviderName = deps.agent.tts?.provider === "elevenlabs" ? "elevenlabs" : deps.agent.tts?.provider === "openai" ? "openai_tts" : deps.agent.elevenlabsKey ? "elevenlabs" : void 0;
+    const sttProviderName = deps.agent.stt ? deps.agent.stt.constructor?.name ?? "custom" : void 0;
+    const ttsProviderName = deps.agent.tts ? deps.agent.tts.constructor?.name ?? "custom" : void 0;
     const providerMode = deps.agent.provider ?? "openai_realtime";
     this.metricsAcc = new CallMetricsAccumulator({
       callId: "",
@@ -2708,7 +2351,7 @@ var StreamHandler = class {
       ttsProvider: ttsProviderName,
       pricing: deps.pricing
     });
-    getLogger().info(`WebSocket connection opened (${deps.bridge.label})`);
+    getLogger().debug(`WebSocket connection opened (${deps.bridge.label})`);
   }
   // ---------------------------------------------------------------------------
   // Public: called by the provider-specific parsers in server.ts
@@ -2724,9 +2367,12 @@ var StreamHandler = class {
     this.metricsAcc.callId = callId;
     if (customParams.caller && !this.caller) this.caller = customParams.caller;
     if (customParams.callee && !this.callee) this.callee = customParams.callee;
-    getLogger().info(`Call started: ${callId}`);
+    const mode = this.deps.agent.engine ? `engine=${this.deps.agent.engine.kind ?? "unknown"}` : "pipeline";
+    getLogger().info(
+      `Call started: ${callId} (${this.deps.bridge.label}, ${mode}, ${sanitizeLogValue(this.caller || "?")} \u2192 ${sanitizeLogValue(this.callee || "?")})`
+    );
     if (Object.keys(customParams).length > 0) {
-      getLogger().info(`Custom params: ${sanitizeLogValue(JSON.stringify(customParams))}`);
+      getLogger().debug(`Custom params: ${sanitizeLogValue(JSON.stringify(customParams))}`);
     }
     this.deps.metricsStore.recordCallStart({
       call_id: callId,
@@ -2774,7 +2420,7 @@ var StreamHandler = class {
             }
           });
           if (recResp.ok) {
-            getLogger().info(`Recording started for ${callId}`);
+            getLogger().debug(`Recording started for ${callId}`);
           } else {
             getLogger().warn(`could not start recording: ${await recResp.text()}`);
           }
@@ -2829,7 +2475,7 @@ var StreamHandler = class {
   }
   /** Handle a DTMF keypress event (Twilio only). */
   async handleDtmf(digit) {
-    getLogger().info(`DTMF: ${digit}`);
+    getLogger().debug(`DTMF: ${digit}`);
     if (this.adapter instanceof OpenAIRealtimeAdapter) {
       await this.adapter.sendText(`The user pressed key ${digit} on their phone keypad.`);
     }
@@ -2888,26 +2534,17 @@ var StreamHandler = class {
   // ---------------------------------------------------------------------------
   async initPipeline(resolvedPrompt) {
     const label = this.deps.bridge.label;
-    this.stt = this.deps.bridge.createStt(this.deps.agent);
-    if (this.deps.agent.tts) {
-      if (this.deps.agent.tts.provider === "elevenlabs") {
-        this.tts = new ElevenLabsTTS(this.deps.agent.tts.apiKey, this.deps.agent.tts.voice ?? "21m00Tcm4TlvDq8ikWAM");
-      }
-      if (this.deps.agent.tts.provider === "openai") {
-        this.tts = new OpenAITTS(this.deps.agent.tts.apiKey, this.deps.agent.tts.voice ?? "alloy");
-      }
-    } else if (this.deps.agent.elevenlabsKey) {
-      this.tts = new ElevenLabsTTS(this.deps.agent.elevenlabsKey, this.deps.agent.voice || "rachel");
-    }
+    this.stt = await this.deps.bridge.createStt(this.deps.agent);
+    this.tts = await createTTS(this.deps.agent);
     if (!this.stt) {
-      getLogger().info(`Pipeline mode (${label}): no STT configured`);
+      getLogger().debug(`Pipeline mode (${label}): no STT configured`);
     }
     if (!this.tts) {
-      getLogger().info(`Pipeline mode (${label}): no TTS configured`);
+      getLogger().debug(`Pipeline mode (${label}): no TTS configured`);
     }
     try {
       if (this.stt) await this.stt.connect();
-      getLogger().info(`Pipeline mode (${label}): STT + TTS connected`);
+      getLogger().debug(`Pipeline mode (${label}): STT + TTS connected`);
     } catch (e) {
       getLogger().error(`Pipeline connect FAILED (${label}):`, e);
       try {
@@ -2940,7 +2577,24 @@ var StreamHandler = class {
         this.history.push({ role: "assistant", text: this.deps.agent.firstMessage, timestamp: Date.now() });
       }
     }
-    if (!this.deps.onMessage && this.deps.config.openaiKey) {
+    if (this.deps.agent.llm) {
+      if (this.deps.onMessage) {
+        throw new Error(
+          "Cannot pass both agent({ llm }) and serve({ onMessage }). Pick one \u2014 `llm` for built-in LLMs, `onMessage` for custom logic."
+        );
+      }
+      this.llmLoop = new LLMLoop(
+        "",
+        // apiKey unused when llmProvider is supplied
+        "",
+        // model unused when llmProvider is supplied
+        resolvedPrompt,
+        this.deps.agent.tools,
+        this.deps.agent.llm
+      );
+      const llmLabel = this.deps.agent.llm.constructor?.name ?? "custom";
+      getLogger().debug(`Built-in LLM loop active (pipeline, ${label}, llm=${llmLabel})`);
+    } else if (!this.deps.onMessage && this.deps.config.openaiKey) {
       let llmModel = this.deps.agent.model || "gpt-4o-mini";
       if (llmModel.includes("realtime")) llmModel = "gpt-4o-mini";
       this.llmLoop = new LLMLoop(
@@ -2949,7 +2603,7 @@ var StreamHandler = class {
         resolvedPrompt,
         this.deps.agent.tools
       );
-      getLogger().info(`Built-in LLM loop active (pipeline, ${label})`);
+      getLogger().debug(`Built-in LLM loop active (pipeline, ${label})`);
     }
     if (this.stt) {
       this.stt.onTranscript(async (transcript) => {
@@ -3010,7 +2664,7 @@ var StreamHandler = class {
   }
   async processTranscript(transcript) {
     if (transcript.text && this.isSpeaking) {
-      getLogger().info(
+      getLogger().debug(
         `Barge-in: caller spoke over agent (${sanitizeLogValue(transcript.text.slice(0, 40))})`
       );
       this.isSpeaking = false;
@@ -3045,17 +2699,17 @@ var StreamHandler = class {
       "cool"
     ]);
     if (HALLUCINATIONS.has(stripped) || stripped === "") {
-      getLogger().info(`Dropped likely STT hallucination: ${sanitizeLogValue(normalised.slice(0, 40))}`);
+      getLogger().debug(`Dropped likely STT hallucination: ${sanitizeLogValue(normalised.slice(0, 40))}`);
       return;
     }
     if (sinceLastMs < 2e3 && normalised === this.lastCommitText) {
-      getLogger().info(
+      getLogger().debug(
         `Dropped duplicate final transcript (${(sinceLastMs / 1e3).toFixed(1)}s since last): ${sanitizeLogValue(normalised.slice(0, 40))}`
       );
       return;
     }
     if (sinceLastMs < 500) {
-      getLogger().info(
+      getLogger().debug(
         `Dropped back-to-back final transcript (${(sinceLastMs / 1e3).toFixed(2)}s since last): ${sanitizeLogValue(normalised.slice(0, 40))}`
       );
       return;
@@ -3063,7 +2717,7 @@ var StreamHandler = class {
     this.lastCommitText = normalised;
     this.lastCommitAt = now;
     const label = this.deps.bridge.label;
-    getLogger().info(`User (${label} pipeline): ${sanitizeLogValue(transcript.text)}`);
+    getLogger().debug(`User (${label} pipeline): ${sanitizeLogValue(transcript.text)}`);
     this.metricsAcc.startTurn();
     this.metricsAcc.recordSttComplete(transcript.text);
     if (this.deps.onTranscript) {
@@ -3078,7 +2732,7 @@ var StreamHandler = class {
     const hookCtx = this.buildHookContext();
     const filteredTranscript = await hookExecutor.runAfterTranscribe(transcript.text, hookCtx);
     if (filteredTranscript === null) {
-      getLogger().info(`afterTranscribe hook vetoed turn (${label})`);
+      getLogger().debug(`afterTranscribe hook vetoed turn (${label})`);
       this.metricsAcc.recordTurnInterrupted();
       return;
     }
@@ -3166,7 +2820,7 @@ var StreamHandler = class {
     if (!this.llmLoop) {
       const guard = checkGuardrails(responseText, this.deps.agent.guardrails);
       if (guard) {
-        getLogger().info(`Guardrail '${guard.name}' triggered (pipeline)`);
+        getLogger().debug(`Guardrail '${guard.name}' triggered (pipeline)`);
         responseText = guard.replacement ?? "I'm sorry, I can't respond to that.";
       }
       this.metricsAcc.recordLlmComplete();
@@ -3244,7 +2898,7 @@ var StreamHandler = class {
     this.adapter = this.deps.buildAIAdapter(resolvedPrompt);
     try {
       await this.adapter.connect();
-      getLogger().info(`AI adapter connected (${label})`);
+      getLogger().debug(`AI adapter connected (${label})`);
     } catch (e) {
       getLogger().error(`AI adapter connect FAILED (${label}):`, e);
       try {
@@ -3286,7 +2940,7 @@ var StreamHandler = class {
       this.deps.bridge.sendMark(this.ws, `audio_${this.chunkCount}`, this.streamSid);
     } else if (type === "transcript_input") {
       const inputText = eventData;
-      getLogger().info(`User (${this.deps.bridge.label}): ${sanitizeLogValue(inputText)}`);
+      getLogger().debug(`User (${this.deps.bridge.label}): ${sanitizeLogValue(inputText)}`);
       this.history.push({ role: "user", text: inputText, timestamp: Date.now() });
       this.metricsAcc.startTurn();
       this.currentAgentText = "";
@@ -3304,7 +2958,7 @@ var StreamHandler = class {
       if (outputText) {
         const triggered = checkGuardrails(outputText, this.deps.agent.guardrails);
         if (triggered) {
-          getLogger().info(`Guardrail '${triggered.name}' triggered`);
+          getLogger().debug(`Guardrail '${triggered.name}' triggered`);
           if (this.adapter instanceof OpenAIRealtimeAdapter) {
             this.adapter.cancelResponse();
             await this.adapter.sendText(triggered.replacement ?? "I'm sorry, I can't respond to that.");
@@ -3363,7 +3017,7 @@ var StreamHandler = class {
         await adapter.sendFunctionResult(fc.call_id, JSON.stringify({ error: "Invalid phone number format", status: "rejected" }));
         return;
       }
-      getLogger().info(`Transferring call to ${transferTo}`);
+      getLogger().debug(`Transferring call to ${transferTo}`);
       await adapter.sendFunctionResult(fc.call_id, JSON.stringify({ status: "transferring", to: transferTo }));
       await this.deps.bridge.transferCall(this.callId, transferTo);
       if (this.deps.onTranscript) {
@@ -3379,7 +3033,7 @@ var StreamHandler = class {
         endArgs = {};
       }
       const reason = endArgs.reason ?? "conversation_complete";
-      getLogger().info(`Ending call (${this.deps.bridge.label}): ${reason}`);
+      getLogger().debug(`Ending call (${this.deps.bridge.label}): ${reason}`);
       await adapter.sendFunctionResult(fc.call_id, JSON.stringify({ status: "ending", reason }));
       await this.deps.bridge.endCall(this.callId, this.ws);
       if (this.deps.onTranscript) {
@@ -3416,10 +3070,11 @@ var StreamHandler = class {
       this.maxDurationTimer = null;
     }
     await this.deps.bridge.queryTelephonyCost(this.metricsAcc, this.callId);
-    const deepgramKey = this.deps.agent.deepgramKey;
-    const deepgramRequestId = this.stt?.requestId;
-    if (deepgramKey && deepgramRequestId) {
-      await queryDeepgramCost(this.metricsAcc, deepgramKey, deepgramRequestId);
+    if (this.stt instanceof DeepgramSTT && this.stt.requestId) {
+      const dgKey = this.stt.apiKey;
+      if (dgKey) {
+        await queryDeepgramCost(this.metricsAcc, dgKey, this.stt.requestId);
+      }
     }
     const finalMetrics = this.metricsAcc.endCall();
     const callEndData = {
@@ -3430,6 +3085,11 @@ var StreamHandler = class {
       transcript: [...this.history.entries],
       metrics: finalMetrics
     };
+    const cost = finalMetrics.cost?.total ?? 0;
+    const latencyP95 = finalMetrics.latency_p95?.total_ms ?? 0;
+    getLogger().info(
+      `Call ended: ${this.callId} (${finalMetrics.duration_seconds.toFixed(1)}s, ${finalMetrics.turns.length} turns, cost=$${cost.toFixed(4)}, p95=${Math.round(latencyP95)}ms)`
+    );
     this.deps.metricsStore.recordCallEnd(
       callEndData,
       finalMetrics
@@ -3466,7 +3126,7 @@ async function queryDeepgramCost(metricsAcc, deepgramKey, deepgramRequestId) {
           const usd = reqData.response?.details?.usd;
           if (usd != null) {
             metricsAcc.setActualSttCost(usd);
-            getLogger().info(`Deepgram actual cost: $${usd}`);
+            getLogger().debug(`Deepgram actual cost: $${usd}`);
           }
         }
       }
@@ -3576,11 +3236,16 @@ function resolveVariables(template, variables) {
   return result;
 }
 function buildAIAdapter(config, agent, resolvedPrompt) {
+  const engine = agent.engine;
   if (agent.provider === "elevenlabs_convai") {
-    const key = agent.elevenlabsKey ?? "";
+    if (!engine || engine.kind !== "elevenlabs_convai") {
+      throw new Error(
+        "ElevenLabs ConvAI mode requires `agent.engine = new ElevenLabsConvAI({...})`."
+      );
+    }
     return new ElevenLabsConvAIAdapter(
-      key,
-      agent.elevenlabsAgentId ?? "",
+      engine.apiKey,
+      engine.agentId,
       agent.voice ?? "21m00Tcm4TlvDq8ikWAM",
       "eleven_turbo_v2_5",
       agent.language ?? "en",
@@ -3593,33 +3258,15 @@ function buildAIAdapter(config, agent, resolvedPrompt) {
     parameters: t.parameters
   })) ?? [];
   const tools = [...agentTools, TRANSFER_CALL_TOOL, END_CALL_TOOL];
+  const openaiKey = engine && engine.kind === "openai_realtime" ? engine.apiKey : config.openaiKey ?? "";
   return new OpenAIRealtimeAdapter(
-    config.openaiKey ?? "",
+    openaiKey,
     agent.model,
     agent.voice,
     resolvedPrompt ?? agent.systemPrompt,
     tools
   );
 }
-function extractDeepgramOptions(options) {
-  if (!options) return {};
-  const get = (snake, camel) => options[snake] ?? options[camel];
-  const out = {};
-  const model = get("model", "model");
-  if (typeof model === "string") out.model = model;
-  const endpointing = get("endpointing_ms", "endpointingMs");
-  if (typeof endpointing === "number") out.endpointingMs = endpointing;
-  const utteranceEnd = get("utterance_end_ms", "utteranceEndMs");
-  if (utteranceEnd === null) out.utteranceEndMs = null;
-  else if (typeof utteranceEnd === "number") out.utteranceEndMs = utteranceEnd;
-  const smart = get("smart_format", "smartFormat");
-  if (typeof smart === "boolean") out.smartFormat = smart;
-  const interim = get("interim_results", "interimResults");
-  if (typeof interim === "boolean") out.interimResults = interim;
-  const vad = get("vad_events", "vadEvents");
-  if (typeof vad === "boolean") out.vadEvents = vad;
-  return out;
-}
 var TwilioBridge = class {
   constructor(config) {
     this.config = config;
@@ -3671,24 +3318,7 @@ var TwilioBridge = class {
     }
   }
   createStt(agent) {
-    const isPipeline = agent.provider === "pipeline";
-    if (agent.stt) {
-      if (agent.stt.provider === "deepgram") {
-        const dgOptions = extractDeepgramOptions(agent.stt.options);
-        if (isPipeline) {
-          return new DeepgramSTT(agent.stt.apiKey, agent.stt.language ?? "en", dgOptions.model, "linear16", 16e3, dgOptions);
-        }
-        return DeepgramSTT.forTwilio(agent.stt.apiKey, agent.stt.language ?? "en", dgOptions.model, dgOptions);
-      } else if (agent.stt.provider === "whisper") {
-        return isPipeline ? new WhisperSTT(agent.stt.apiKey, "whisper-1", agent.stt.language ?? "en") : WhisperSTT.forTwilio(agent.stt.apiKey, agent.stt.language ?? "en");
-      }
-    } else if (agent.deepgramKey) {
-      if (isPipeline) {
-        return new DeepgramSTT(agent.deepgramKey, agent.language ?? "en", "nova-3", "linear16", 16e3);
-      }
-      return DeepgramSTT.forTwilio(agent.deepgramKey, agent.language ?? "en");
-    }
-    return null;
+    return createSTT(agent);
   }
   async queryTelephonyCost(metricsAcc, callId) {
     if (this.config.twilioSid && this.config.twilioToken && callId) {
@@ -3835,24 +3465,7 @@ var TelnyxBridge = class {
     ws.close();
   }
   createStt(agent) {
-    if (agent.stt) {
-      if (agent.stt.provider === "deepgram") {
-        const dgOptions = extractDeepgramOptions(agent.stt.options);
-        return new DeepgramSTT(
-          agent.stt.apiKey,
-          agent.stt.language ?? "en",
-          dgOptions.model ?? "nova-3",
-          "linear16",
-          16e3,
-          dgOptions
-        );
-      } else if (agent.stt.provider === "whisper") {
-        return new WhisperSTT(agent.stt.apiKey, "whisper-1", agent.stt.language ?? "en");
-      }
-    } else if (agent.deepgramKey) {
-      return new DeepgramSTT(agent.deepgramKey, agent.language ?? "en", "nova-3", "linear16", 16e3);
-    }
-    return null;
+    return createSTT(agent);
   }
   async queryTelephonyCost(metricsAcc, callId) {
     if (this.config.telnyxKey && callId) {
@@ -4790,8 +4403,6 @@ var TestSession = class {
 export {
   OpenAIRealtimeAdapter,
   ElevenLabsConvAIAdapter,
-  DeepgramSTT,
-  WhisperSTT,
   DEFAULT_PRICING,
   mergePricing,
   calculateSttCost,
@@ -4807,8 +4418,7 @@ export {
   RemoteMessageHandler,
   isRemoteUrl,
   isWebSocketUrl,
-  ElevenLabsTTS,
-  OpenAITTS,
+  DeepgramSTT,
   CallMetricsAccumulator,
   mulawToPcm16,
   pcm16ToMulaw,