getpatter 0.5.1 → 0.5.3

package/dist/{chunk-B6C3KIBG.mjs → chunk-FIFIWBL7.mjs}

@@ -1,12 +1,15 @@
 import {
   getLogger
-} from "./chunk-FMNRCP5X.mjs";
+} from "./chunk-VJVDG4V5.mjs";
+import {
+  __require
+} from "./chunk-QHHBUCMT.mjs";
 
 // src/test-mode.ts
 import { createInterface } from "readline";
 
 // src/server.ts
-import crypto3 from "crypto";
+import crypto4 from "crypto";
 import express from "express";
 import { createServer } from "http";
 import { WebSocketServer } from "ws";
@@ -14,14 +17,24 @@ import { WebSocketServer } from "ws";
 // src/providers/openai-realtime.ts
 import WebSocket from "ws";
 var OpenAIRealtimeAdapter = class {
-  constructor(apiKey, model = "gpt-4o-mini-realtime-preview", voice = "alloy", instructions = "", tools) {
+  constructor(apiKey, model = "gpt-realtime-mini", voice = "alloy", instructions = "", tools, audioFormat = "g711_ulaw", options = {}) {
     this.apiKey = apiKey;
     this.model = model;
     this.voice = voice;
     this.instructions = instructions;
     this.tools = tools;
+    this.audioFormat = audioFormat;
+    this.options = options;
   }
   ws = null;
+  eventCallbacks = /* @__PURE__ */ new Set();
+  messageListenerAttached = false;
+  heartbeat = null;
+  // Track the in-flight assistant item id so we can truncate cleanly on
+  // barge-in (see ``cancelResponse``) — matches the Python adapter.
+  currentResponseItemId = null;
+  currentResponseAudioMs = 0;
+  options;
   async connect() {
     const url = `wss://api.openai.com/v1/realtime?model=${encodeURIComponent(this.model)}`;
     this.ws = new WebSocket(url, {
@@ -45,13 +58,24 @@ var OpenAIRealtimeAdapter = class {
         if (msg.type === "session.created" && !sessionCreated) {
           sessionCreated = true;
           const config = {
-            input_audio_format: "g711_ulaw",
-            output_audio_format: "g711_ulaw",
+            input_audio_format: this.audioFormat,
+            output_audio_format: this.audioFormat,
             voice: this.voice,
             instructions: this.instructions || "You are a helpful voice assistant. Be concise.",
-            turn_detection: { type: "server_vad", threshold: 0.5, prefix_padding_ms: 300, silence_duration_ms: 500 },
-            input_audio_transcription: { model: "whisper-1" }
+            turn_detection: {
+              type: this.options.vadType ?? "server_vad",
+              threshold: 0.5,
+              prefix_padding_ms: 300,
+              silence_duration_ms: this.options.silenceDurationMs ?? 300
+            },
+            input_audio_transcription: { model: this.options.inputAudioTranscriptionModel ?? "whisper-1" }
           };
+          if (this.options.temperature !== void 0) config.temperature = this.options.temperature;
+          if (this.options.maxResponseOutputTokens !== void 0) {
+            config.max_response_output_tokens = this.options.maxResponseOutputTokens;
+          }
+          if (this.options.modalities !== void 0) config.modalities = this.options.modalities;
+          if (this.options.toolChoice !== void 0) config.tool_choice = this.options.toolChoice;
           if (this.tools?.length) {
             config.tools = this.tools.map((t) => ({
               type: "function",
@@ -92,19 +116,45 @@ var OpenAIRealtimeAdapter = class {
       ws.on("message", onSetupMessage);
       ws.on("error", onSetupError);
     });
+    this.heartbeat = setInterval(() => {
+      try {
+        this.ws?.ping();
+      } catch {
+      }
+    }, 2e4);
+    this.ensureMessageListener();
   }
   sendAudio(mulawAudio) {
     if (!this.ws || this.ws.readyState !== WebSocket.OPEN) return;
     this.ws.send(JSON.stringify({ type: "input_audio_buffer.append", audio: mulawAudio.toString("base64") }));
   }
+  /**
+   * Register a listener for parsed realtime events.
+   *
+   * Previously every call attached a new ``ws.on('message')`` handler,
+   * which leaked listeners across retries and multi-consumer hooks. We now
+   * route all traffic through a single persistent handler that fans out to
+   * a Set of callbacks. Use {@link offEvent} to remove one.
+   */
   onEvent(callback) {
-    if (!this.ws) return;
-    const safeInvoke = (type, data) => {
-      void Promise.resolve(callback(type, data)).catch(
-        (err) => getLogger().error("onEvent callback error:", err)
-      );
+    this.eventCallbacks.add(callback);
+    this.ensureMessageListener();
+  }
+  offEvent(callback) {
+    this.eventCallbacks.delete(callback);
+  }
+  ensureMessageListener() {
+    if (this.messageListenerAttached || !this.ws) return;
+    this.messageListenerAttached = true;
+    const ws = this.ws;
+    const dispatch = (type, payload) => {
+      for (const cb of this.eventCallbacks) {
+        void Promise.resolve(cb(type, payload)).catch(
+          (err) => getLogger().error("onEvent callback error:", err)
+        );
+      }
     };
-    this.ws.on("message", (raw) => {
+    ws.on("message", (raw) => {
       let data;
       try {
         data = JSON.parse(raw.toString());
@@ -114,24 +164,61 @@ var OpenAIRealtimeAdapter = class {
       }
       const t = data.type;
       if (t === "response.audio.delta") {
-        safeInvoke("audio", Buffer.from(data.delta ?? "", "base64"));
+        const buf = Buffer.from(data.delta ?? "", "base64");
+        this.currentResponseAudioMs += estimateAudioMs(buf, this.audioFormat);
+        dispatch("audio", buf);
       } else if (t === "response.audio_transcript.delta") {
-        safeInvoke("transcript_output", data.delta);
+        dispatch("transcript_output", data.delta);
+      } else if (t === "response.content_part.added" || t === "response.output_item.added") {
+        const itemId = data.item?.id ?? data.item_id ?? null;
+        if (itemId) {
+          this.currentResponseItemId = itemId;
+          this.currentResponseAudioMs = 0;
+        }
       } else if (t === "input_audio_buffer.speech_started") {
-        safeInvoke("speech_started", null);
+        dispatch("speech_started", null);
+      } else if (t === "input_audio_buffer.speech_stopped") {
+        dispatch("speech_stopped", null);
       } else if (t === "conversation.item.input_audio_transcription.completed") {
-        safeInvoke("transcript_input", data.transcript);
+        dispatch("transcript_input", data.transcript);
       } else if (t === "response.function_call_arguments.done") {
-        safeInvoke("function_call", { call_id: data.call_id, name: data.name, arguments: data.arguments });
+        dispatch("function_call", { call_id: data.call_id, name: data.name, arguments: data.arguments });
       } else if (t === "response.done") {
-        safeInvoke("response_done", data.response ?? null);
+        this.currentResponseItemId = null;
+        this.currentResponseAudioMs = 0;
+        dispatch("response_done", data.response ?? null);
       } else if (t === "error") {
-        safeInvoke("error", data.error);
+        dispatch("error", data.error);
+      }
+    });
+    ws.on("close", (code, reason) => {
+      if (code !== 1e3) {
+        dispatch("error", {
+          type: "connection_closed",
+          code,
+          reason: reason?.toString() ?? ""
+        });
       }
     });
+    ws.on("error", (err) => {
+      dispatch("error", { type: "socket_error", message: err?.message ?? String(err) });
+    });
   }
   cancelResponse() {
-    this.ws?.send(JSON.stringify({ type: "response.cancel" }));
+    if (!this.ws) return;
+    if (this.currentResponseItemId) {
+      try {
+        this.ws.send(JSON.stringify({
+          type: "conversation.item.truncate",
+          item_id: this.currentResponseItemId,
+          content_index: 0,
+          audio_end_ms: this.currentResponseAudioMs
+        }));
+      } catch (err) {
+        getLogger().debug?.(`conversation.item.truncate failed: ${String(err)}`);
+      }
+    }
+    this.ws.send(JSON.stringify({ type: "response.cancel" }));
   }
   async sendText(text) {
     this.ws?.send(JSON.stringify({
@@ -148,28 +235,148 @@ var OpenAIRealtimeAdapter = class {
     this.ws?.send(JSON.stringify({ type: "response.create" }));
   }
   close() {
+    if (this.heartbeat) {
+      clearInterval(this.heartbeat);
+      this.heartbeat = null;
+    }
+    this.eventCallbacks.clear();
+    this.messageListenerAttached = false;
     this.ws?.close();
     this.ws = null;
   }
 };
+function estimateAudioMs(chunk, format) {
+  if (chunk.length === 0) return 0;
+  if (format === "g711_ulaw" || format === "g711_alaw") return Math.floor(chunk.length / 8);
+  if (format === "pcm16") {
+    return Math.floor(chunk.length / 48);
+  }
+  return 0;
+}
 
 // src/providers/elevenlabs-convai.ts
 import WebSocket2 from "ws";
 var ELEVENLABS_CONVAI_URL = "wss://api.elevenlabs.io/v1/convai/conversation";
-var ElevenLabsConvAIAdapter = class {
-  constructor(apiKey, agentId = "", voiceId = "21m00Tcm4TlvDq8ikWAM", _modelId = "eleven_turbo_v2_5", _language = "en", firstMessage = "") {
-    this.apiKey = apiKey;
-    this.agentId = agentId;
-    this.voiceId = voiceId;
-    this.firstMessage = firstMessage;
-  }
+var ELEVENLABS_SIGNED_URL = "https://api.elevenlabs.io/v1/convai/conversation/get-signed-url";
+var AGENT_SILENCE_MS = 500;
+var ElevenLabsConvAIAdapter = class _ElevenLabsConvAIAdapter {
   ws = null;
   eventCallback = null;
-  async connect() {
-    const url = this.agentId ? `${ELEVENLABS_CONVAI_URL}?agent_id=${encodeURIComponent(this.agentId)}` : ELEVENLABS_CONVAI_URL;
-    this.ws = new WebSocket2(url, {
-      headers: { "xi-api-key": this.apiKey }
+  apiKey;
+  agentId;
+  voiceId;
+  // Exposed for parity with Python SDK (`self.model_id`). ConvAI does not
+  // accept a client-side model override today, but we preserve the value so
+  // callers can introspect it and we can ship the override the day the
+  // server exposes it.
+  modelId;
+  language;
+  firstMessage;
+  // Exposed publicly so the stream handler can detect μ-law negotiation
+  // (``"ulaw_8000"``) and skip resampling / transcoding on the audio path.
+  outputAudioFormat;
+  inputAudioFormat;
+  useSignedUrl;
+  // Populated from `conversation_initiation_metadata`.
+  conversationId = null;
+  agentOutputAudioFormat = null;
+  userInputAudioFormat = null;
+  agentSpeaking = false;
+  silenceTimer = null;
+  closePromise = null;
+  constructor(apiKeyOrOptions, agentId = "", voiceId = "EXAVITQu4vr4xnSDxMaL", firstMessage = "") {
+    if (typeof apiKeyOrOptions === "object") {
+      const o = apiKeyOrOptions;
+      this.apiKey = o.apiKey;
+      this.agentId = o.agentId ?? "";
+      this.voiceId = o.voiceId ?? "EXAVITQu4vr4xnSDxMaL";
+      this.modelId = o.modelId ?? "eleven_flash_v2_5";
+      this.language = o.language ?? "it";
+      this.firstMessage = o.firstMessage ?? "";
+      this.outputAudioFormat = o.outputAudioFormat;
+      this.inputAudioFormat = o.inputAudioFormat;
+      this.useSignedUrl = o.useSignedUrl ?? false;
+    } else {
+      this.apiKey = apiKeyOrOptions;
+      this.agentId = agentId;
+      this.voiceId = voiceId;
+      this.modelId = "eleven_flash_v2_5";
+      this.language = "it";
+      this.firstMessage = firstMessage;
+      this.outputAudioFormat = void 0;
+      this.inputAudioFormat = void 0;
+      this.useSignedUrl = false;
+    }
+  }
+  // ------------------------------------------------------------------
+  // Telephony factories
+  // ------------------------------------------------------------------
+  /**
+   * Build an adapter pre-configured for Twilio Media Streams.
+   *
+   * Negotiates `ulaw_8000` for both `outputAudioFormat` and
+   * `inputAudioFormat`, matching Twilio's μ-law @ 8 kHz wire format. The
+   * SDK's stream handler detects this and skips the 8 kHz → 16 kHz inbound
+   * resample and the 16 kHz → 8 kHz / PCM → μ-law outbound transcode.
+   * Saves ~30–80 ms first-byte plus per-frame CPU on every turn.
+   */
+  static forTwilio(apiKey, agentId, options = {}) {
+    return new _ElevenLabsConvAIAdapter({
+      ...options,
+      apiKey,
+      agentId,
+      outputAudioFormat: "ulaw_8000",
+      inputAudioFormat: "ulaw_8000"
+    });
+  }
+  /**
+   * Build an adapter pre-configured for Telnyx bidirectional media.
+   *
+   * Telnyx negotiates PCMU @ 8 kHz when `streaming_start` sets
+   * `stream_bidirectional_codec=PCMU` (the SDK default). Picking
+   * `ulaw_8000` on both ConvAI directions removes every transcode on the
+   * audio path — same optimization as `forTwilio`.
+   */
+  static forTelnyx(apiKey, agentId, options = {}) {
+    return new _ElevenLabsConvAIAdapter({
+      ...options,
+      apiKey,
+      agentId,
+      outputAudioFormat: "ulaw_8000",
+      inputAudioFormat: "ulaw_8000"
     });
+  }
+  async fetchSignedUrl() {
+    if (!this.agentId) {
+      throw new Error("useSignedUrl=true requires agentId");
+    }
+    const url = `${ELEVENLABS_SIGNED_URL}?agent_id=${encodeURIComponent(this.agentId)}`;
+    const resp = await fetch(url, {
+      method: "GET",
+      headers: { "xi-api-key": this.apiKey },
+      signal: AbortSignal.timeout(15e3)
+    });
+    if (!resp.ok) {
+      const body = await resp.text();
+      throw new Error(`ElevenLabs signed-url error ${resp.status}: ${body}`);
+    }
+    const data = await resp.json();
+    if (!data.signed_url) {
+      throw new Error("ElevenLabs signed-url response missing 'signed_url'");
+    }
+    return data.signed_url;
+  }
+  async connect() {
+    let wsUrl;
+    let wsOptions;
+    if (this.useSignedUrl) {
+      wsUrl = await this.fetchSignedUrl();
+      wsOptions = void 0;
+    } else {
+      wsUrl = this.agentId ? `${ELEVENLABS_CONVAI_URL}?agent_id=${encodeURIComponent(this.agentId)}` : ELEVENLABS_CONVAI_URL;
+      wsOptions = { headers: { "xi-api-key": this.apiKey } };
+    }
+    this.ws = new WebSocket2(wsUrl, wsOptions);
     await new Promise((resolve, reject) => {
       const timeout = setTimeout(
         () => reject(new Error("ElevenLabs ConvAI connect timeout")),
@@ -177,17 +384,22 @@ var ElevenLabsConvAIAdapter = class {
       );
       this.ws.once("open", () => {
         clearTimeout(timeout);
+        const agentCfg = {};
+        if (this.firstMessage) agentCfg["first_message"] = this.firstMessage;
+        if (this.language) agentCfg["language"] = this.language;
+        const override = {
+          tts: this.outputAudioFormat ? { voice_id: this.voiceId, output_format: this.outputAudioFormat } : { voice_id: this.voiceId }
+        };
+        if (this.inputAudioFormat) {
+          override["asr"] = { input_format: this.inputAudioFormat };
+        }
+        if (Object.keys(agentCfg).length > 0) {
+          override["agent"] = agentCfg;
+        }
         const config = {
           type: "conversation_initiation_client_data",
-          conversation_config_override: {
-            tts: { voice_id: this.voiceId }
-          }
+          conversation_config_override: override
         };
-        if (this.firstMessage) {
-          config["conversation_config_override"]["agent"] = {
-            first_message: this.firstMessage
-          };
-        }
         this.ws.send(JSON.stringify(config));
         resolve();
       });
@@ -196,54 +408,176 @@ var ElevenLabsConvAIAdapter = class {
         reject(err);
       });
     });
+    this.ws.on("error", (err) => {
+      getLogger().error("ElevenLabs ConvAI WS error:", err);
+      this.safeInvoke("error", err instanceof Error ? err.message : String(err));
+    });
+    this.ws.on("close", (code, reason) => {
+      this.clearSilenceTimer();
+      this.safeInvoke("close", {
+        code,
+        reason: reason?.toString() ?? ""
+      });
+    });
     this.ws.on("message", (raw) => {
-      const cb = this.eventCallback;
-      if (!cb) return;
-      const safeInvoke = (type, data) => {
-        void Promise.resolve(cb(type, data)).catch(
-          (err) => getLogger().error("onEvent callback error:", err)
-        );
-      };
       let parsed;
       try {
         parsed = JSON.parse(raw.toString());
       } catch {
         return;
       }
-      const msgType = parsed["type"];
-      if (msgType === "audio") {
-        const audioB64 = parsed["audio"];
-        if (audioB64) {
-          safeInvoke("audio", Buffer.from(audioB64, "base64"));
-        }
-      } else if (msgType === "user_transcript") {
-        safeInvoke("transcript_input", parsed["text"] ?? "");
-      } else if (msgType === "agent_response") {
-        safeInvoke("transcript_output", parsed["text"] ?? "");
-        safeInvoke("response_done", null);
-      } else if (msgType === "interruption") {
-        safeInvoke("interruption", null);
-      } else if (msgType === "error") {
-        safeInvoke("error", parsed);
-      }
+      this.handleMessage(parsed);
     });
   }
+  safeInvoke(type, data) {
+    const cb = this.eventCallback;
+    if (!cb) return;
+    void Promise.resolve(cb(type, data)).catch(
+      (err) => getLogger().error("onEvent callback error:", err)
+    );
+  }
+  respondToPing(eventId, delayMs) {
+    const send = () => {
+      if (!this.ws || this.ws.readyState !== WebSocket2.OPEN) return;
+      try {
+        this.ws.send(JSON.stringify({ type: "pong", event_id: eventId }));
+      } catch (err) {
+        getLogger().warn("ElevenLabs ConvAI pong send failed:", err);
+      }
+    };
+    if (delayMs && delayMs > 0) {
+      setTimeout(send, delayMs);
+    } else {
+      send();
+    }
+  }
+  clearSilenceTimer() {
+    if (this.silenceTimer) {
+      clearTimeout(this.silenceTimer);
+      this.silenceTimer = null;
+    }
+  }
+  finalizeAgentTurn() {
+    this.clearSilenceTimer();
+    if (this.agentSpeaking) {
+      this.agentSpeaking = false;
+      this.safeInvoke("response_done", null);
+    }
+  }
+  scheduleSilenceDone() {
+    this.clearSilenceTimer();
+    this.silenceTimer = setTimeout(() => {
+      if (this.agentSpeaking) {
+        this.agentSpeaking = false;
+        this.safeInvoke("response_done", null);
+      }
+    }, AGENT_SILENCE_MS);
+  }
+  handleMessage(parsed) {
+    const msgType = parsed["type"];
+    if (msgType === "ping") {
+      const pingPayload = parsed["ping_event"] ?? parsed["ping"] ?? {};
+      const eventId = pingPayload["event_id"] ?? parsed["event_id"];
+      const pingMs = pingPayload["ping_ms"] ?? 0;
+      this.respondToPing(eventId, pingMs);
+      return;
+    }
+    if (msgType === "conversation_initiation_metadata") {
+      const meta = parsed["conversation_initiation_metadata_event"] ?? parsed;
+      this.conversationId = meta["conversation_id"] ?? this.conversationId;
+      this.agentOutputAudioFormat = meta["agent_output_audio_format"] ?? this.agentOutputAudioFormat;
+      this.userInputAudioFormat = meta["user_input_audio_format"] ?? this.userInputAudioFormat;
+      this.finalizeAgentTurn();
+      return;
+    }
+    if (msgType === "audio") {
+      const audioEvt = parsed["audio_event"];
+      let audioB64;
+      if (audioEvt) {
+        audioB64 = audioEvt["audio_base_64"] ?? audioEvt["audio"];
+      }
+      if (!audioB64) {
+        audioB64 = parsed["audio"];
+      }
+      if (audioB64) {
+        this.agentSpeaking = true;
+        this.safeInvoke("audio", Buffer.from(audioB64, "base64"));
+        this.scheduleSilenceDone();
+      }
+      return;
+    }
+    if (msgType === "user_transcript") {
+      const evt = parsed["user_transcription_event"] ?? parsed;
+      const text = evt["user_transcript"] ?? evt["text"] ?? "";
+      this.finalizeAgentTurn();
+      this.safeInvoke("transcript_input", text);
+      return;
+    }
+    if (msgType === "agent_response") {
+      const evt = parsed["agent_response_event"] ?? parsed;
+      const text = evt["agent_response"] ?? evt["text"] ?? "";
+      this.safeInvoke("transcript_output", text);
+      this.agentSpeaking = true;
+      this.safeInvoke("response_start", { text });
+      return;
+    }
+    if (msgType === "interruption") {
+      this.finalizeAgentTurn();
+      this.safeInvoke("interruption", null);
+      return;
+    }
+    if (msgType === "error") {
+      const errText = parsed["message"] ?? parsed["error"] ?? JSON.stringify(parsed);
+      getLogger().error("ElevenLabs ConvAI error:", errText);
+      this.safeInvoke("error", errText);
+      return;
+    }
+  }
   sendAudio(audioBytes) {
     if (!this.ws || this.ws.readyState !== WebSocket2.OPEN) return;
     this.ws.send(
       JSON.stringify({
-        type: "audio",
-        audio: audioBytes.toString("base64")
+        user_audio_chunk: audioBytes.toString("base64")
       })
     );
   }
   onEvent(callback) {
     this.eventCallback = callback;
   }
-  close() {
-    this.ws?.close();
-    this.ws = null;
-    this.eventCallback = null;
+  async close() {
+    this.clearSilenceTimer();
+    if (!this.ws) {
+      this.eventCallback = null;
+      return;
+    }
+    if (this.closePromise) {
+      await this.closePromise;
+      return;
+    }
+    const ws = this.ws;
+    this.closePromise = new Promise((resolve) => {
+      if (ws.readyState === WebSocket2.CLOSED || ws.readyState === WebSocket2.CLOSING) {
+        resolve();
+        return;
+      }
+      const done = () => {
+        resolve();
+      };
+      ws.once("close", done);
+      ws.once("error", done);
+      try {
+        ws.close();
+      } catch {
+        resolve();
+      }
+    });
+    try {
+      await this.closePromise;
+    } finally {
+      this.ws = null;
+      this.eventCallback = null;
+      this.closePromise = null;
+    }
   }
 };
 
@@ -258,21 +592,57 @@ async function createTTS(agent) {
 // src/pricing.ts
 var DEFAULT_PRICING = {
   // STT — per minute of audio processed
-  deepgram: { unit: "minute", price: 43e-4 },
+  // Deepgram Nova-3 streaming (monolingual) — the default model Patter ships.
+  // The previous $0.0043/min was the batch rate; streaming is $0.0077/min per
+  // deepgram.com/pricing. For multilingual Nova-3 ($0.0092/min) override.
+  deepgram: { unit: "minute", price: 77e-4 },
   whisper: { unit: "minute", price: 6e-3 },
-  // TTS — per 1,000 characters synthesized
-  elevenlabs: { unit: "1k_chars", price: 0.18 },
+  // AssemblyAI Universal-Streaming — $0.15/hr = $0.0025/min
+  assemblyai: { unit: "minute", price: 25e-4 },
+  // Cartesia ink-whisper streaming STT — ~$0.15/hr on usage plans
+  cartesia_stt: { unit: "minute", price: 25e-4 },
+  // Soniox real-time STT — $0.12/hr = $0.002/min
+  soniox: { unit: "minute", price: 2e-3 },
+  // Speechmatics Pro tier — $0.24/hr = $0.0040/min (new users land here).
+  // Previous $0.0173 default reflected a legacy Standard tier that was
+  // retired; users were being over-billed ~4.3x.
+  speechmatics: { unit: "minute", price: 4e-3 },
+  // TTS — per 1,000 characters synthesized.
+  // ElevenLabs default model is eleven_flash_v2_5 billed at $0.06/1k via the
+  // direct API. The previous $0.18 matched only the Creator plan overage.
+  elevenlabs: { unit: "1k_chars", price: 0.06 },
   openai_tts: { unit: "1k_chars", price: 0.015 },
-  // OpenAI Realtime — per token
+  openai_tts_hd: { unit: "1k_chars", price: 0.03 },
+  // Cartesia Sonic TTS — ~1 credit/char, effective $0.030/1k chars on usage plans
+  cartesia_tts: { unit: "1k_chars", price: 0.03 },
+  // Rime mist v2 — $0.030/1k chars pay-as-you-go
+  rime: { unit: "1k_chars", price: 0.03 },
+  // LMNT aurora/blizzard — $0.050/1k chars Indie overage
+  lmnt: { unit: "1k_chars", price: 0.05 },
+  // OpenAI Realtime — per token.
+  // Calibrated for gpt-4o-mini-realtime-preview (the Patter default):
+  //   audio  input  $10 / M  ->  0.00001 per token
+  //   audio  output $20 / M  ->  0.00002 per token
+  //   text   input  $0.60/ M ->  0.0000006 per token
+  //   text   output $2.40/ M ->  0.0000024 per token
+  // For gpt-4o-realtime-preview multiply by ~10, for gpt-realtime by ~3.
   openai_realtime: {
     unit: "token",
-    audio_input_per_token: 1e-4,
-    audio_output_per_token: 4e-4,
-    text_input_per_token: 5e-6,
-    text_output_per_token: 2e-5
+    audio_input_per_token: 1e-5,
+    audio_output_per_token: 2e-5,
+    text_input_per_token: 6e-7,
+    text_output_per_token: 24e-7,
+    // Prompt caching rates (official): audio cached $0.30/M ~= 3% of full,
+    // text cached $0.06/M = 10% of full. OpenAI bills the cached portion of
+    // input_token_details.audio_tokens / text_tokens at these reduced rates.
+    cached_audio_input_per_token: 3e-7,
+    cached_text_input_per_token: 6e-8
   },
-  // Telephony — per minute of call duration
-  twilio: { unit: "minute", price: 0.013 },
+  // Telephony — per minute of call duration.
+  // twilio default = US inbound local (the 99% case for voice agents receiving
+  // calls on a local number). For US toll-free inbound ($0.022/min) or US
+  // outbound local ($0.0140/min), override via Patter({ pricing: { twilio: {...} } }).
+  twilio: { unit: "minute", price: 85e-4 },
   telnyx: { unit: "minute", price: 7e-3 }
 };
 function mergePricing(overrides) {
@@ -281,22 +651,22 @@ function mergePricing(overrides) {
     merged[k] = { ...v };
   }
   if (!overrides) return merged;
-  for (const [provider, values] of Object.entries(overrides)) {
-    if (merged[provider]) {
-      merged[provider] = { ...merged[provider], ...values };
+  for (const [provider2, values] of Object.entries(overrides)) {
+    if (merged[provider2]) {
+      merged[provider2] = { ...merged[provider2], ...values };
     } else {
-      merged[provider] = { unit: "minute", ...values };
+      merged[provider2] = { ...values };
     }
   }
   return merged;
 }
-function calculateSttCost(provider, audioSeconds, pricing) {
-  const config = pricing[provider];
+function calculateSttCost(provider2, audioSeconds, pricing) {
+  const config = pricing[provider2];
   if (!config || config.unit !== "minute") return 0;
   return audioSeconds / 60 * (config.price ?? 0);
 }
-function calculateTtsCost(provider, characterCount, pricing) {
-  const config = pricing[provider];
+function calculateTtsCost(provider2, characterCount, pricing) {
+  const config = pricing[provider2];
   if (!config || config.unit !== "1k_chars") return 0;
   return characterCount / 1e3 * (config.price ?? 0);
 }
@@ -305,21 +675,126 @@ function calculateRealtimeCost(usage, pricing) {
   if (!config || config.unit !== "token") return 0;
   const input = usage.input_token_details ?? {};
   const output = usage.output_token_details ?? {};
+  const cachedAudioRate = config.cached_audio_input_per_token ?? config.audio_input_per_token ?? 0;
+  const cachedTextRate = config.cached_text_input_per_token ?? config.text_input_per_token ?? 0;
+  const totalAudioIn = input.audio_tokens ?? 0;
+  const totalTextIn = input.text_tokens ?? 0;
+  let cachedAudioIn;
+  let cachedTextIn;
+  const details = input.cached_tokens_details;
+  if (details && (details.audio_tokens !== void 0 || details.text_tokens !== void 0)) {
+    cachedAudioIn = Math.min(details.audio_tokens ?? 0, totalAudioIn);
+    cachedTextIn = Math.min(details.text_tokens ?? 0, totalTextIn);
+  } else if (input.cached_tokens && input.cached_tokens > 0) {
+    const totalIn = totalAudioIn + totalTextIn;
+    const ratio = totalIn > 0 ? input.cached_tokens / totalIn : 0;
+    cachedAudioIn = Math.min(Math.round(totalAudioIn * ratio), totalAudioIn);
+    cachedTextIn = Math.min(Math.round(totalTextIn * ratio), totalTextIn);
+  } else {
+    cachedAudioIn = 0;
+    cachedTextIn = 0;
+  }
   let cost = 0;
-  cost += (input.audio_tokens ?? 0) * (config.audio_input_per_token ?? 0);
-  cost += (input.text_tokens ?? 0) * (config.text_input_per_token ?? 0);
+  cost += (totalAudioIn - cachedAudioIn) * (config.audio_input_per_token ?? 0);
+  cost += cachedAudioIn * cachedAudioRate;
+  cost += (totalTextIn - cachedTextIn) * (config.text_input_per_token ?? 0);
+  cost += cachedTextIn * cachedTextRate;
   cost += (output.audio_tokens ?? 0) * (config.audio_output_per_token ?? 0);
   cost += (output.text_tokens ?? 0) * (config.text_output_per_token ?? 0);
-  return cost;
+  return Math.max(0, cost);
+}
+function calculateRealtimeCachedSavings(usage, pricing) {
+  const config = pricing.openai_realtime;
+  if (!config || config.unit !== "token") return 0;
+  const input = usage.input_token_details ?? {};
+  const cached = input.cached_tokens_details ?? {};
+  const cachedAudioRate = config.cached_audio_input_per_token ?? config.audio_input_per_token ?? 0;
+  const cachedTextRate = config.cached_text_input_per_token ?? config.text_input_per_token ?? 0;
+  const cachedAudio = Math.min(cached.audio_tokens ?? 0, input.audio_tokens ?? 0);
+  const cachedText = Math.min(cached.text_tokens ?? 0, input.text_tokens ?? 0);
+  const fullAudio = cachedAudio * (config.audio_input_per_token ?? 0);
+  const fullText = cachedText * (config.text_input_per_token ?? 0);
+  const discountedAudio = cachedAudio * cachedAudioRate;
+  const discountedText = cachedText * cachedTextRate;
+  return Math.max(0, fullAudio + fullText - (discountedAudio + discountedText));
+}
+var llmPricing = {
+  anthropic: {
+    "claude-opus-4-7": {
+      input: 15,
+      output: 75,
+      cache_read: 1.5,
+      cache_write: 18.75
+    },
+    "claude-sonnet-4-6": {
+      input: 3,
+      output: 15,
+      cache_read: 0.3,
+      cache_write: 3.75
+    },
+    "claude-haiku-4-5": {
+      input: 1,
+      output: 5,
+      cache_read: 0.1,
+      cache_write: 1.25
+    }
+  },
+  google: {
+    "gemini-2.5-pro": { input: 1.25, output: 10 },
+    "gemini-2.5-flash": { input: 0.3, output: 2.5 },
+    "gemini-live-2.5-flash-native-audio": { input: 0.3, output: 2.5 }
+  },
+  groq: {
+    "llama-3.3-70b-versatile": { input: 0.59, output: 0.79 },
+    "llama-3.1-8b-instant": { input: 0.05, output: 0.08 }
+  },
+  cerebras: {
+    "llama-3.3-70b": { input: 0.85, output: 1.2 },
+    "qwen-3-32b": { input: 0.4, output: 0.8 }
+  },
+  // OpenAI Chat Completions (non-Realtime) — mirrors sdk-py pricing table.
+  // Rates are per 1M tokens (USD), cache_read = cached input rate.
+  openai: {
+    "gpt-4o": { input: 2.5, output: 10, cache_read: 1.25 },
+    "gpt-4o-mini": { input: 0.15, output: 0.6, cache_read: 0.075 },
+    "gpt-4.1": { input: 3, output: 12, cache_read: 0.75 },
+    "gpt-4.1-mini": { input: 0.8, output: 3.2, cache_read: 0.2 },
+    "o3": { input: 2, output: 8, cache_read: 0.5 },
+    "o4-mini": { input: 1.1, output: 4.4, cache_read: 0.275 }
+  }
+};
+function calculateLlmCost(provider2, model, inputTokens, outputTokens, cacheReadTokens = 0, cacheWriteTokens = 0) {
+  const providerTable = llmPricing[provider2];
+  if (!providerTable) return 0;
+  let rates = providerTable[model];
+  if (!rates) {
+    let bestKey = "";
+    for (const key of Object.keys(providerTable)) {
+      if (model.startsWith(key) && key.length > bestKey.length) {
+        bestKey = key;
+      }
+    }
+    if (bestKey) rates = providerTable[bestKey];
+  }
+  if (!rates) return 0;
+  let cost = 0;
+  cost += inputTokens / 1e6 * (rates.input ?? 0);
+  cost += outputTokens / 1e6 * (rates.output ?? 0);
+  cost += cacheReadTokens / 1e6 * (rates.cache_read ?? 0);
+  cost += cacheWriteTokens / 1e6 * (rates.cache_write ?? 0);
+  return Math.max(0, cost);
 }
-function calculateTelephonyCost(provider, durationSeconds, pricing) {
-  const config = pricing[provider];
+function calculateTelephonyCost(provider2, durationSeconds, pricing) {
+  const config = pricing[provider2];
   if (!config || config.unit !== "minute") return 0;
-  return durationSeconds / 60 * (config.price ?? 0);
+  const minutes = provider2 === "twilio" ? Math.ceil(durationSeconds / 60) : durationSeconds / 60;
+  return minutes * (config.price ?? 0);
 }
 
 // src/dashboard/store.ts
 import { EventEmitter } from "events";
+import * as fs from "fs";
+import * as path from "path";
 var MetricsStore = class extends EventEmitter {
   maxCalls;
   calls = [];
@@ -482,6 +957,10 @@ var MetricsStore = class extends EventEmitter {
     }
     return null;
   }
+  /** Look up an active call by id (returns undefined if not active or unknown). */
+  getActive(callId) {
+    return this.activeCalls.get(callId);
+  }
   getActiveCalls() {
     return Array.from(this.activeCalls.values());
   }
@@ -547,7 +1026,102 @@ var MetricsStore = class extends EventEmitter {
   get callCount() {
     return this.calls.length;
   }
+  /**
+   * Rebuild the in-memory call list from `metadata.json` files written by
+   * `CallLogger` under `<logRoot>/calls/YYYY/MM/DD/<call_id>/`. Idempotent:
+   * call_ids already in the store are skipped. Errors per file are logged
+   * and swallowed so a single corrupt entry doesn't block hydration.
+   *
+   * Returns the number of calls newly added to the store.
+   *
+   * Safe to call before any traffic; intended to run once at server startup.
+   */
+  hydrate(logRoot) {
+    if (!logRoot) return 0;
+    const callsRoot = path.join(logRoot, "calls");
+    if (!fs.existsSync(callsRoot)) return 0;
+    const collected = [];
+    const seen = new Set(this.calls.map((c) => c.call_id));
+    const walk = (dir, depth) => {
+      let entries;
+      try {
+        entries = fs.readdirSync(dir, { withFileTypes: true });
+      } catch {
+        return;
+      }
+      for (const entry of entries) {
+        const childPath = path.join(dir, entry.name);
+        if (depth < 3) {
+          if (entry.isDirectory() && /^\d+$/.test(entry.name)) {
+            walk(childPath, depth + 1);
+          }
+          continue;
+        }
+        if (!entry.isDirectory()) continue;
+        const metadataPath = path.join(childPath, "metadata.json");
+        if (!fs.existsSync(metadataPath)) continue;
+        try {
+          const raw = fs.readFileSync(metadataPath, "utf8");
+          const meta = JSON.parse(raw);
+          const callId = meta.call_id || entry.name;
+          if (!callId || seen.has(callId)) continue;
+          const record = metadataToCallRecord(callId, meta);
+          if (record === null) {
+            getLogger().debug(
+              `MetricsStore.hydrate: skipping ${metadataPath}: unparseable started_at`
+            );
+            continue;
+          }
+          collected.push(record);
+          seen.add(callId);
+        } catch (err) {
+          getLogger().debug(
+            `MetricsStore.hydrate: skipping ${metadataPath}: ${String(err)}`
+          );
+        }
+      }
+    };
+    walk(callsRoot, 0);
+    collected.sort((a, b) => (a.started_at || 0) - (b.started_at || 0));
+    for (const rec of collected) {
+      if (this.calls.some((c) => c.call_id === rec.call_id)) continue;
+      this.calls.push(rec);
+      if (this.calls.length > this.maxCalls) {
+        this.calls = this.calls.slice(-this.maxCalls);
+      }
+    }
+    return collected.length;
+  }
 };
+function metadataToCallRecord(callId, meta) {
+  const startedAt = parseTimestamp(meta.started_at);
+  if (startedAt === null) return null;
+  const endedAt = parseTimestamp(meta.ended_at);
+  const status = meta.status || "completed";
+  const metrics = meta.metrics && typeof meta.metrics === "object" ? meta.metrics : null;
+  const transcript = Array.isArray(meta.transcript) ? meta.transcript : [];
+  return {
+    call_id: callId,
+    caller: meta.caller || "",
+    callee: meta.callee || "",
+    direction: meta.direction || "inbound",
+    started_at: startedAt,
+    ended_at: endedAt ?? void 0,
+    status,
+    metrics,
+    transcript
+  };
+}
+function parseTimestamp(raw) {
+  if (typeof raw === "number") {
+    return Number.isFinite(raw) ? raw : null;
+  }
+  if (typeof raw === "string") {
+    const ms = Date.parse(raw);
+    return Number.isFinite(ms) ? ms / 1e3 : null;
+  }
+  return null;
+}
 
 // src/dashboard/auth.ts
 import crypto from "crypto";
@@ -1628,10 +2202,52 @@ function isWebSocketUrl(url) {
 
 // src/providers/deepgram-stt.ts
 import WebSocket3 from "ws";
+
+// src/errors.ts
+var PatterError = class extends Error {
+  constructor(message) {
+    super(message);
+    this.name = "PatterError";
+  }
+};
+var PatterConnectionError = class extends PatterError {
+  constructor(message) {
+    super(message);
+    this.name = "PatterConnectionError";
+  }
+};
+var AuthenticationError = class extends PatterError {
+  constructor(message) {
+    super(message);
+    this.name = "AuthenticationError";
+  }
+};
+var ProvisionError = class extends PatterError {
+  constructor(message) {
+    super(message);
+    this.name = "ProvisionError";
+  }
+};
+var RateLimitError = class extends PatterConnectionError {
+  constructor(message) {
+    super(message);
+    this.name = "RateLimitError";
+  }
+};
+
+// src/providers/deepgram-stt.ts
 var DEEPGRAM_WS_URL = "wss://api.deepgram.com/v1/listen";
+var KEEPALIVE_INTERVAL_MS = 4e3;
+var FINALIZE_DRAIN_MS = 100;
+var CLOSE_LATENCY_BUDGET_MS = 500;
+var RECONNECT_CLOSE_CODES = /* @__PURE__ */ new Set([1006, 1011]);
 var DeepgramSTT = class _DeepgramSTT {
   ws = null;
-  callbacks = [];
+  transcriptCallbacks = /* @__PURE__ */ new Set();
+  errorCallbacks = /* @__PURE__ */ new Set();
+  keepaliveTimer = null;
+  running = false;
+  reconnectAttempted = false;
   /** Request ID from Deepgram — used to query actual cost post-call. */
   requestId = "";
   apiKey;
@@ -1653,7 +2269,7 @@ var DeepgramSTT = class _DeepgramSTT {
     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.smartFormat = opts.smartFormat ?? false;
     this.interimResults = opts.interimResults ?? true;
     this.vadEvents = opts.vadEvents ?? true;
   }
@@ -1661,7 +2277,7 @@ var DeepgramSTT = class _DeepgramSTT {
   static forTwilio(apiKey, language = "en", model = "nova-3", options = {}) {
     return new _DeepgramSTT(apiKey, language, model, "mulaw", 8e3, options);
   }
-  async connect() {
+  buildUrl() {
     const params = new URLSearchParams({
       model: this.model,
       language: this.language,
@@ -1677,69 +2293,201 @@ var DeepgramSTT = class _DeepgramSTT {
     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, {
+    return `${DEEPGRAM_WS_URL}?${params.toString()}`;
+  }
+  async connect() {
+    await this.openSocket();
+    this.running = true;
+    this.reconnectAttempted = false;
+  }
+  async openSocket() {
+    const url = this.buildUrl();
+    const ws = new WebSocket3(url, {
       headers: { Authorization: `Token ${this.apiKey}` }
     });
+    this.ws = ws;
     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) => {
+      let settled = false;
+      const settle = (fn) => {
+        if (settled) return;
+        settled = true;
         clearTimeout(timer);
-        reject(err);
+        fn();
+      };
+      const timer = setTimeout(
+        () => settle(() => reject(new PatterConnectionError("Deepgram connect timeout"))),
+        1e4
+      );
+      ws.once("open", () => settle(resolve));
+      ws.once("error", (err) => settle(() => reject(err)));
+      ws.once("unexpected-response", (_req, res) => {
+        const status = res?.statusCode ?? 0;
+        settle(() => {
+          if (status === 401 || status === 403) {
+            reject(new AuthenticationError(`Deepgram rejected the API key (HTTP ${status}).`));
+            return;
+          }
+          if (status === 429) {
+            reject(new RateLimitError("Deepgram rate limit exceeded (HTTP 429)."));
+            return;
+          }
+          reject(new PatterConnectionError(`Deepgram WebSocket upgrade failed (HTTP ${status}).`));
+        });
       });
     });
-    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;
+    ws.on("message", (raw) => this.handleMessage(raw.toString()));
+    ws.on("close", (code, reason) => this.handleClose(code, reason.toString()));
+    ws.on("error", (err) => this.handleError(err));
+    this.keepaliveTimer = setInterval(() => {
+      if (this.ws && this.ws.readyState === WebSocket3.OPEN) {
+        try {
+          this.ws.send(JSON.stringify({ type: "KeepAlive" }));
+        } catch {
+        }
       }
-      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) {
+    }, KEEPALIVE_INTERVAL_MS);
+  }
+  clearKeepalive() {
+    if (this.keepaliveTimer) {
+      clearInterval(this.keepaliveTimer);
+      this.keepaliveTimer = null;
+    }
+  }
+  handleMessage(raw) {
+    let data;
+    try {
+      data = JSON.parse(raw);
+    } catch {
+      return;
+    }
+    if (data.type === "Metadata" && data.request_id) {
+      this.requestId = data.request_id;
+      return;
+    }
+    if (data.type === "SpeechStarted") {
+      this.emitTranscript({
+        text: "",
+        isFinal: false,
+        confidence: 0,
+        eventType: "SpeechStarted",
+        requestId: this.requestId || void 0
+      });
+      return;
+    }
+    if (data.type === "UtteranceEnd") {
+      this.emitTranscript({
+        text: "",
+        isFinal: true,
+        confidence: 0,
+        eventType: "UtteranceEnd",
+        requestId: this.requestId || void 0
+      });
+      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 speechFinal = Boolean(data.speech_final);
+    const transcript = {
+      text,
+      isFinal: Boolean(data.is_final) || speechFinal,
+      confidence: best.confidence ?? 0,
+      speechFinal,
+      fromFinalize: Boolean(data.from_finalize),
+      requestId: this.requestId || void 0,
+      words: best.words,
+      eventType: "Results"
+    };
+    this.emitTranscript(transcript);
+  }
+  emitTranscript(transcript) {
+    for (const cb of this.transcriptCallbacks) {
+      try {
         cb(transcript);
+      } catch (err) {
+        getLogger().error(`DeepgramSTT transcript callback threw: ${String(err)}`);
       }
-    });
+    }
+  }
+  emitError(err) {
+    for (const cb of this.errorCallbacks) {
+      try {
+        cb(err);
+      } catch (cbErr) {
+        getLogger().error(`DeepgramSTT error callback threw: ${String(cbErr)}`);
+      }
+    }
+  }
+  handleError(err) {
+    getLogger().error(`DeepgramSTT WebSocket error: ${err.message}`);
+    this.emitError(err);
+  }
+  handleClose(code, reason) {
+    this.clearKeepalive();
+    if (!this.running) {
+      return;
+    }
+    const closeError = new PatterConnectionError(
+      `Deepgram WebSocket closed (code=${code}${reason ? `, reason=${reason}` : ""}).`
+    );
+    this.emitError(closeError);
+    if (RECONNECT_CLOSE_CODES.has(code) && !this.reconnectAttempted) {
+      this.reconnectAttempted = true;
+      this.openSocket().catch((err) => {
+        this.running = false;
+        this.emitError(err instanceof Error ? err : new Error(String(err)));
+      });
+    } else {
+      this.running = false;
+    }
   }
   sendAudio(audio) {
     if (!this.ws || this.ws.readyState !== WebSocket3.OPEN) return;
+    if (audio.length === 0) 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);
+    this.transcriptCallbacks.add(callback);
+  }
+  offTranscript(callback) {
+    this.transcriptCallbacks.delete(callback);
+  }
+  onError(callback) {
+    this.errorCallbacks.add(callback);
+  }
+  offError(callback) {
+    this.errorCallbacks.delete(callback);
   }
   close() {
-    if (this.ws) {
-      try {
-        this.ws.send(JSON.stringify({ type: "CloseStream" }));
-      } catch {
+    this.running = false;
+    this.clearKeepalive();
+    const ws = this.ws;
+    if (!ws) return;
+    this.ws = null;
+    const sendSafe = (payload) => {
+      if (ws.readyState === WebSocket3.OPEN) {
+        try {
+          ws.send(payload);
+        } catch {
+        }
       }
-      this.ws.close();
-      this.ws = null;
+    };
+    const finishClose = () => {
+      sendSafe(JSON.stringify({ type: "CloseStream" }));
+      try {
+        ws.close();
+      } catch {
+      }
+    };
+    if (ws.readyState !== WebSocket3.OPEN) {
+      finishClose();
+      return;
     }
+    sendSafe(JSON.stringify({ type: "Finalize" }));
+    setTimeout(finishClose, Math.min(FINALIZE_DRAIN_MS, CLOSE_LATENCY_BUDGET_MS));
   }
 };
 
@@ -1752,11 +2500,16 @@ function hrTimeMs() {
   const [sec, ns] = process.hrtime();
   return sec * 1e3 + ns / 1e6;
 }
-function p95(values) {
+function percentile(values, p) {
   if (values.length === 0) return 0;
   const sorted = [...values].sort((a, b) => a - b);
-  const idx = Math.min(Math.floor(sorted.length * 0.95), sorted.length - 1);
-  return sorted[idx];
+  if (sorted.length === 1) return sorted[0];
+  const rank = p * (sorted.length - 1);
+  const lo = Math.floor(rank);
+  const hi = Math.ceil(rank);
+  if (lo === hi) return sorted[lo];
+  const frac = rank - lo;
+  return sorted[lo] + (sorted[hi] - sorted[lo]) * frac;
 }
 var CallMetricsAccumulator = class {
   callId;
@@ -1771,19 +2524,54 @@ var CallMetricsAccumulator = class {
   // Per-turn timing state
   _turnStart = null;
   _sttComplete = null;
+  _llmFirstToken = null;
+  _llmFirstSentenceComplete = null;
   _llmComplete = null;
   _ttsFirstByte = null;
+  /** Last TTS audio byte sent (hrTimeMs). Stamped by ``recordTtsComplete`` /
+   *  ``recordTtsCompleteTs``. Used to compute ``tts_total_ms``. */
+  _ttsLastByte = null;
+  /** Endpoint signal (hrTimeMs) — VAD stop or STT speech_final, whichever
+   *  fires first. Used to compute ``endpoint_ms``. */
+  _endpointSignalAt = null;
+  /** Monotonic stamp of LLM dispatch (paired with ``_endpointSignalAt``). */
+  _turnCommittedMono = null;
+  /** Barge-in detected timestamp (hrTimeMs). */
+  _bargeinDetectedAt = null;
+  /** TTS-stopped timestamp after barge-in (hrTimeMs). */
+  _bargeinStoppedAt = null;
   _turnUserText = "";
   _turnSttAudioSeconds = 0;
   // Cumulative usage counters
   _totalSttAudioSeconds = 0;
   _totalTtsCharacters = 0;
   _totalRealtimeCost = 0;
+  _totalRealtimeCachedSavings = 0;
   _sttByteCount = 0;
   _sttSampleRate = 16e3;
   _sttBytesPerSample = 2;
   _actualTelephonyCost = null;
   _actualSttCost = null;
+  // Fix 10: accumulated LLM token cost for non-Realtime pipeline mode.
+  _totalLlmCost = 0;
+  // ---- EventBus integration (item 3) ----
+  _eventBus;
+  // ---- EOUMetrics — 4 timestamps (item 4) ----
+  /** Timestamp (hrTimeMs) when VAD emitted speech_end. */
+  _vadStoppedAt = null;
+  /** Timestamp (hrTimeMs) when STT emitted its final transcript. */
+  _sttFinalAt = null;
+  /** Timestamp (hrTimeMs) when the transcript was committed to the LLM. */
+  _turnCommittedAt = null;
+  /** Delta (ms) from turn-committed to on_user_turn_completed hook done. */
+  _onUserTurnCompletedDelayMs = null;
+  // ---- InterruptionMetrics — simplified no-ML (item 5) ----
+  _numInterruptions = 0;
+  _numBackchannels = 0;
+  _overlapStartedAt = null;
+  // ---- report_only_initial_ttfb (item 6) ----
+  _reportOnlyInitialTtfb;
+  _initialTtfbEmitted = false;
   constructor(opts) {
     this.callId = opts.callId;
     this.providerMode = opts.providerMode;
@@ -1793,6 +2581,15 @@ var CallMetricsAccumulator = class {
     this.llmProvider = opts.llmProvider ?? "";
     this._pricing = mergePricing(opts.pricing);
     this._callStart = hrTimeMs();
+    this._eventBus = opts.eventBus;
+    this._reportOnlyInitialTtfb = opts.reportOnlyInitialTtfb ?? false;
+  }
+  /**
+   * Attach (or replace) an EventBus after construction.
+   * Useful when the bus is created after the accumulator (e.g. in tests).
+   */
+  attachEventBus(bus) {
+    this._eventBus = bus;
   }
   /** Configure audio format for STT byte-to-seconds conversion. */
   configureSttFormat(sampleRate = 16e3, bytesPerSample = 2) {
@@ -1807,17 +2604,60 @@ var CallMetricsAccumulator = class {
   startTurn() {
     this._turnStart = hrTimeMs();
     this._sttComplete = null;
+    this._llmFirstToken = null;
+    this._llmFirstSentenceComplete = null;
     this._llmComplete = null;
     this._ttsFirstByte = null;
+    this._ttsLastByte = null;
+    this._endpointSignalAt = null;
+    this._turnCommittedMono = null;
+    this._bargeinDetectedAt = null;
+    this._bargeinStoppedAt = null;
     this._turnUserText = "";
     this._turnSttAudioSeconds = 0;
+    this._vadStoppedAt = null;
+    this._sttFinalAt = null;
+    this._turnCommittedAt = null;
+    this._onUserTurnCompletedDelayMs = null;
+    this._eventBus?.emit("turn_started", { callId: this.callId });
+  }
+  /**
+   * Start a new turn only if no turn is currently open.
+   * Use this at inbound-audio ingestion points so the turn timer begins
+   * on the first audio byte rather than just before recordSttComplete().
+   */
+  startTurnIfIdle() {
+    if (this._turnStart === null) {
+      this.startTurn();
+    }
   }
   recordSttComplete(text, audioSeconds = 0) {
     this._sttComplete = hrTimeMs();
+    this._sttFinalAt = this._sttComplete;
+    if (this._endpointSignalAt === null) {
+      this._endpointSignalAt = this._sttComplete;
+    }
     this._turnUserText = text;
     this._turnSttAudioSeconds = audioSeconds;
     this._totalSttAudioSeconds += audioSeconds;
   }
+  /** Record the timestamp of the first LLM token (TTFT). No-op after first call. */
+  recordLlmFirstToken() {
+    if (this._llmFirstToken === null) {
+      this._llmFirstToken = hrTimeMs();
+    }
+  }
+  /**
+   * Record when the sentence chunker emits the first complete sentence.
+   * Used as the TTS span start so tts_ms reflects true TTS-provider latency
+   * rather than the gap from llm_complete (which fires after the full response).
+   * No-op after first call.
+   */
+  recordLlmFirstSentenceComplete() {
+    if (this._llmFirstSentenceComplete === null) {
+      this._llmFirstSentenceComplete = hrTimeMs();
+    }
+  }
   recordLlmComplete() {
     this._llmComplete = hrTimeMs();
   }
@@ -1825,9 +2665,40 @@ var CallMetricsAccumulator = class {
     if (this._ttsFirstByte === null) {
       this._ttsFirstByte = hrTimeMs();
     }
+    if (this._reportOnlyInitialTtfb && this._initialTtfbEmitted) {
+      return;
+    }
+    this._initialTtfbEmitted = true;
   }
   recordTtsComplete(text) {
     this._totalTtsCharacters += text.length;
+    if (this._ttsLastByte === null) {
+      this._ttsLastByte = hrTimeMs();
+    }
+  }
+  /**
+   * Capture the timestamp when the last TTS audio byte was sent on the wire.
+   * Useful when the caller wants to record the timing without bumping the
+   * character counter (e.g. interrupted turns where audio actually went out
+   * but synthesis was truncated).
+   */
+  recordTtsCompleteTs(ts) {
+    this._ttsLastByte = ts ?? hrTimeMs();
+  }
+  /**
+   * Mark the moment a user interrupt (barge-in) was detected. Pairs with
+   * ``recordTtsStopped`` to compute ``bargein_ms``.
+   */
+  recordBargeinDetected(ts) {
+    this._bargeinDetectedAt = ts ?? hrTimeMs();
+  }
+  /**
+   * Mark the moment TTS playback was actually halted after a barge-in. Call
+   * this *after* ``sendClear`` returns. Pairs with ``recordBargeinDetected``
+   * to compute ``bargein_ms``.
+   */
+  recordTtsStopped(ts) {
+    this._bargeinStoppedAt = ts ?? hrTimeMs();
   }
   recordTurnComplete(agentText) {
     const latency = this._computeTurnLatency();
@@ -1842,6 +2713,8 @@ var CallMetricsAccumulator = class {
     };
     this._turns.push(turn);
     this._resetTurnState();
+    this._eventBus?.emit("turn_ended", { callId: this.callId, turn });
+    this._eventBus?.emit("metrics_collected", { callId: this.callId, turn });
     return turn;
   }
   recordTurnInterrupted() {
@@ -1860,12 +2733,111 @@ var CallMetricsAccumulator = class {
     this._resetTurnState();
     return turn;
   }
+  // ---- EOU metrics (item 4) ----
+  /**
+   * Record the moment VAD emitted speech_end for the current utterance.
+   * @param ts Optional override timestamp in hrTimeMs units (defaults to now).
+   */
+  recordVadStop(ts) {
+    this._vadStoppedAt = ts ?? hrTimeMs();
+    if (this._endpointSignalAt === null) {
+      this._endpointSignalAt = this._vadStoppedAt;
+    }
+  }
+  /**
+   * Record the moment the STT provider delivered its final transcript.
+   * Aliased to the same instant as recordSttComplete() when called from
+   * the standard pipeline; can be called independently for custom pipelines.
+   * @param ts Optional override timestamp in hrTimeMs units.
+   */
+  recordSttFinalTimestamp(ts) {
+    this._sttFinalAt = ts ?? hrTimeMs();
+    if (this._endpointSignalAt === null) {
+      this._endpointSignalAt = this._sttFinalAt;
+    }
+  }
+  /**
+   * Record the moment the transcript was committed to the LLM (turn start).
+   * After this call, ``emitEouMetrics()`` can produce a complete EOUMetrics payload.
+   * @param ts Optional override timestamp in hrTimeMs units.
+   */
+  recordTurnCommitted(ts) {
+    this._turnCommittedAt = ts ?? hrTimeMs();
+    this._turnCommittedMono = hrTimeMs();
+    this.emitEouMetrics();
+  }
+  /**
+   * Record the delta (ms) between turn-committed and when on_user_turn_completed
+   * pipeline hook finished.  Stored for inclusion in the next ``emitEouMetrics``
+   * call (or an explicit re-emit if desired).
+   */
+  recordOnUserTurnCompletedDelay(delayMs) {
+    this._onUserTurnCompletedDelayMs = delayMs;
+  }
+  /**
+   * Compute and emit EOUMetrics when all three prerequisite timestamps are
+   * available (VAD stop, STT final, turn committed).
+   *
+   * ``endOfUtteranceDelay``     = sttFinal − vadStopped  (ms)
+   * ``transcriptionDelay``       = turnCommitted − vadStopped  (ms)
+   * ``onUserTurnCompletedDelay`` = caller-supplied delta (ms) or 0
+   */
+  emitEouMetrics() {
+    if (this._vadStoppedAt === null || this._sttFinalAt === null || this._turnCommittedAt === null) {
+      return;
+    }
+    const payload = {
+      timestamp: Date.now() / 1e3,
+      endOfUtteranceDelay: Math.max(0, this._sttFinalAt - this._vadStoppedAt),
+      transcriptionDelay: Math.max(0, this._turnCommittedAt - this._vadStoppedAt),
+      onUserTurnCompletedDelay: this._onUserTurnCompletedDelayMs ?? 0
+    };
+    this._eventBus?.emit("eou_metrics", payload);
+  }
+  // ---- InterruptionMetrics (item 5) ----
+  /**
+   * Record that a caller utterance started overlapping with agent speech.
+   * Call this when VAD detects speech_start during TTS playback.
+   * @param ts Optional override timestamp in hrTimeMs units.
+   */
+  recordOverlapStart(ts) {
+    this._overlapStartedAt = ts ?? hrTimeMs();
+  }
+  /**
+   * Record that the overlap ended.  Emits ``InterruptionMetrics`` via the
+   * event bus.
+   *
+   * @param wasInterruption  true → barge-in (increments ``numInterruptions``),
+   *                         false → backchannel (increments ``numBackchannels``).
+   * @param ts Optional override timestamp in hrTimeMs units.
+   */
+  recordOverlapEnd(wasInterruption, ts) {
+    const now = ts ?? hrTimeMs();
+    const detectionDelay = this._overlapStartedAt !== null ? Math.max(0, now - this._overlapStartedAt) : 0;
+    this._overlapStartedAt = null;
+    if (wasInterruption) {
+      this._numInterruptions++;
+    } else {
+      this._numBackchannels++;
+    }
+    const payload = {
+      timestamp: Date.now() / 1e3,
+      // Simplified: totalDuration == detectionDelay (no ML prediction window)
+      totalDuration: detectionDelay,
+      predictionDuration: 0,
+      detectionDelay,
+      numInterruptions: this._numInterruptions,
+      numBackchannels: this._numBackchannels
+    };
+    this._eventBus?.emit("interruption", payload);
+  }
   // ---- Usage tracking ----
   addSttAudioBytes(byteCount) {
     this._sttByteCount += byteCount;
   }
   recordRealtimeUsage(usage) {
     this._totalRealtimeCost += calculateRealtimeCost(usage, this._pricing);
+    this._totalRealtimeCachedSavings += calculateRealtimeCachedSavings(usage, this._pricing);
   }
   setActualTelephonyCost(cost) {
     this._actualTelephonyCost = cost;
@@ -1873,28 +2845,62 @@ var CallMetricsAccumulator = class {
   setActualSttCost(cost) {
     this._actualSttCost = cost;
   }
+  /**
+   * Accumulate LLM token cost for pipeline mode (non-Realtime).
+   *
+   * Called by LLMLoop.run() when a usage chunk arrives from the provider.
+   * Mirrors Python's CallMetricsAccumulator.record_llm_usage().
+   *
+   * @param provider   LLM provider key (e.g. 'openai', 'anthropic')
+   * @param model      Model name (e.g. 'gpt-4o-mini')
+   * @param inputTokens       Total input tokens (includes cached)
+   * @param outputTokens      Total output tokens
+   * @param cacheReadTokens   Cached input tokens (subtracted from input before billing full rate)
+   * @param cacheWriteTokens  Cache write tokens (billed at cache_write rate if present)
+   */
+  recordLlmUsage(provider2, model, inputTokens, outputTokens, cacheReadTokens = 0, cacheWriteTokens = 0) {
+    this._totalLlmCost += calculateLlmCost(
+      provider2,
+      model,
+      inputTokens,
+      outputTokens,
+      cacheReadTokens,
+      cacheWriteTokens
+    );
+  }
   // ---- Finalize ----
   endCall() {
     const duration = (hrTimeMs() - this._callStart) / 1e3;
+    if (this.turnActive) {
+      this.recordTurnInterrupted();
+    }
     if (this._totalSttAudioSeconds === 0 && this._sttByteCount > 0) {
       this._totalSttAudioSeconds = this._sttByteCount / (this._sttSampleRate * this._sttBytesPerSample);
     }
     const cost = this._computeCost(duration);
     const latencyAvg = this._computeAverageLatency();
-    const latencyP95 = this._computeP95Latency();
-    return {
+    const latencyP50 = this._computePercentileLatency(0.5);
+    const latencyP90 = this._computePercentileLatency(0.9);
+    const latencyP95 = this._computePercentileLatency(0.95);
+    const latencyP99 = this._computePercentileLatency(0.99);
+    const metrics = {
       call_id: this.callId,
       duration_seconds: round(duration, 2),
       turns: [...this._turns],
       cost,
       latency_avg: latencyAvg,
+      latency_p50: latencyP50,
+      latency_p90: latencyP90,
       latency_p95: latencyP95,
+      latency_p99: latencyP99,
       provider_mode: this.providerMode,
       stt_provider: this.sttProvider,
       tts_provider: this.ttsProvider,
       llm_provider: this.llmProvider,
       telephony_provider: this.telephonyProvider
     };
+    this._eventBus?.emit("call_ended", { callId: this.callId, metrics });
+    return metrics;
   }
   getCostSoFar() {
     const duration = (hrTimeMs() - this._callStart) / 1e3;
@@ -1904,36 +2910,68 @@ var CallMetricsAccumulator = class {
   _resetTurnState() {
     this._turnStart = null;
     this._sttComplete = null;
+    this._llmFirstToken = null;
+    this._llmFirstSentenceComplete = null;
     this._llmComplete = null;
     this._ttsFirstByte = null;
+    this._ttsLastByte = null;
+    this._endpointSignalAt = null;
+    this._turnCommittedMono = null;
+    this._bargeinDetectedAt = null;
+    this._bargeinStoppedAt = null;
     this._turnUserText = "";
     this._turnSttAudioSeconds = 0;
   }
   _computeTurnLatency() {
     let stt_ms = 0;
     let llm_ms = 0;
+    let llm_ttft_ms;
+    let llm_total_ms;
     let tts_ms = 0;
     let total_ms = 0;
+    let endpoint_ms;
+    let bargein_ms;
+    let tts_total_ms;
     if (this._turnStart !== null && this._sttComplete !== null) {
       stt_ms = this._sttComplete - this._turnStart;
     }
-    if (this._sttComplete !== null && this._llmComplete !== null) {
+    if (this._sttComplete !== null && this._llmFirstToken !== null) {
+      llm_ttft_ms = Math.max(0, this._llmFirstToken - this._sttComplete);
+      llm_ms = llm_ttft_ms;
+    } else if (this._sttComplete !== null && this._llmComplete !== null) {
       llm_ms = this._llmComplete - this._sttComplete;
     }
-    if (this._llmComplete !== null && this._ttsFirstByte !== null) {
-      tts_ms = this._ttsFirstByte - this._llmComplete;
+    if (this._sttComplete !== null && this._llmComplete !== null) {
+      llm_total_ms = Math.max(0, this._llmComplete - this._sttComplete);
+    }
+    const ttsSpanStart = this._llmFirstSentenceComplete ?? this._llmComplete;
+    if (ttsSpanStart !== null && this._ttsFirstByte !== null) {
+      tts_ms = this._ttsFirstByte - ttsSpanStart;
+      if (tts_ms < 0) tts_ms = 0;
     }
     if (this._turnStart !== null && this._ttsFirstByte !== null) {
       total_ms = this._ttsFirstByte - this._turnStart;
     }
-    if (total_ms > 0 && stt_ms === 0 && llm_ms === 0 && tts_ms === 0) {
-      llm_ms = total_ms;
+    if (this._endpointSignalAt !== null && this._turnCommittedMono !== null) {
+      endpoint_ms = Math.max(0, this._turnCommittedMono - this._endpointSignalAt);
+    }
+    if (this._bargeinDetectedAt !== null && this._bargeinStoppedAt !== null) {
+      bargein_ms = Math.max(0, this._bargeinStoppedAt - this._bargeinDetectedAt);
+    }
+    const ttsTotalRef = this._llmFirstToken ?? this._llmFirstSentenceComplete ?? this._llmComplete;
+    if (ttsTotalRef !== null && this._ttsLastByte !== null) {
+      tts_total_ms = Math.max(0, this._ttsLastByte - ttsTotalRef);
     }
     return {
       stt_ms: round(stt_ms, 1),
       llm_ms: round(llm_ms, 1),
+      ...llm_ttft_ms !== void 0 ? { llm_ttft_ms: round(llm_ttft_ms, 1) } : {},
+      ...llm_total_ms !== void 0 ? { llm_total_ms: round(llm_total_ms, 1) } : {},
       tts_ms: round(tts_ms, 1),
-      total_ms: round(total_ms, 1)
+      total_ms: round(total_ms, 1),
+      ...endpoint_ms !== void 0 ? { endpoint_ms: round(endpoint_ms, 1) } : {},
+      ...bargein_ms !== void 0 ? { bargein_ms: round(bargein_ms, 1) } : {},
+      ...tts_total_ms !== void 0 ? { tts_total_ms: round(tts_total_ms, 1) } : {}
     };
   }
   _computeCost(durationSeconds) {
@@ -1951,7 +2989,7 @@ var CallMetricsAccumulator = class {
     } else {
       stt = this._actualSttCost !== null ? this._actualSttCost : calculateSttCost(this.sttProvider, this._totalSttAudioSeconds, this._pricing);
       tts = calculateTtsCost(this.ttsProvider, this._totalTtsCharacters, this._pricing);
-      llm = 0;
+      llm = this._totalLlmCost;
     }
     const telephony = this._actualTelephonyCost !== null ? this._actualTelephonyCost : calculateTelephonyCost(this.telephonyProvider, durationSeconds, this._pricing);
     const total = stt + tts + llm + telephony;
@@ -1960,30 +2998,78 @@ var CallMetricsAccumulator = class {
       tts: round(tts, 6),
       llm: round(llm, 6),
       telephony: round(telephony, 6),
-      total: round(total, 6)
+      total: round(total, 6),
+      // Always emit (default 0) for parity with Python dataclass where
+      // llm_cached_savings is a required field with default 0.0.
+      llm_cached_savings: round(Math.max(0, this._totalRealtimeCachedSavings), 6)
     };
   }
+  /**
+   * Turns eligible for latency statistics.
+   *
+   * Excludes turns marked ``[interrupted]`` (barge-in, cancelled replacements)
+   * because their recorded latency either reflects partial state or zero —
+   * including them would drag every p95/avg bucket toward meaningless numbers.
+   */
+  _completedTurns() {
+    return this._turns.filter(
+      (t) => t.agent_text !== "[interrupted]" && t.latency.total_ms > 0
+    );
+  }
   _computeAverageLatency() {
-    if (this._turns.length === 0) {
+    const turns = this._completedTurns();
+    if (turns.length === 0) {
       return { stt_ms: 0, llm_ms: 0, tts_ms: 0, total_ms: 0 };
     }
-    const n = this._turns.length;
+    const n = turns.length;
+    const ttftValues = turns.map((t) => t.latency.llm_ttft_ms ?? 0).filter((v) => v > 0);
+    const ttftAvg = ttftValues.length > 0 ? round(ttftValues.reduce((s, v) => s + v, 0) / ttftValues.length, 1) : void 0;
+    const optAvg = (key) => {
+      const vals = turns.map((t) => t.latency[key]).filter((v) => typeof v === "number" && v > 0);
+      return vals.length > 0 ? round(vals.reduce((s, v) => s + v, 0) / vals.length, 1) : void 0;
+    };
+    const llmTotalAvg = optAvg("llm_total_ms");
+    const endpointAvg = optAvg("endpoint_ms");
+    const bargeinAvg = optAvg("bargein_ms");
+    const ttsTotalAvg = optAvg("tts_total_ms");
     return {
-      stt_ms: round(this._turns.reduce((s, t) => s + t.latency.stt_ms, 0) / n, 1),
-      llm_ms: round(this._turns.reduce((s, t) => s + t.latency.llm_ms, 0) / n, 1),
-      tts_ms: round(this._turns.reduce((s, t) => s + t.latency.tts_ms, 0) / n, 1),
-      total_ms: round(this._turns.reduce((s, t) => s + t.latency.total_ms, 0) / n, 1)
+      stt_ms: round(turns.reduce((s, t) => s + t.latency.stt_ms, 0) / n, 1),
+      llm_ms: round(turns.reduce((s, t) => s + t.latency.llm_ms, 0) / n, 1),
+      ...ttftAvg !== void 0 ? { llm_ttft_ms: ttftAvg } : {},
+      ...llmTotalAvg !== void 0 ? { llm_total_ms: llmTotalAvg } : {},
+      tts_ms: round(turns.reduce((s, t) => s + t.latency.tts_ms, 0) / n, 1),
+      total_ms: round(turns.reduce((s, t) => s + t.latency.total_ms, 0) / n, 1),
+      ...endpointAvg !== void 0 ? { endpoint_ms: endpointAvg } : {},
+      ...bargeinAvg !== void 0 ? { bargein_ms: bargeinAvg } : {},
+      ...ttsTotalAvg !== void 0 ? { tts_total_ms: ttsTotalAvg } : {}
     };
   }
-  _computeP95Latency() {
-    if (this._turns.length === 0) {
+  _computePercentileLatency(p) {
+    const turns = this._completedTurns();
+    if (turns.length === 0) {
       return { stt_ms: 0, llm_ms: 0, tts_ms: 0, total_ms: 0 };
     }
+    const nonZero = (vals) => vals.filter((v) => v > 0);
+    const ttftSamples = nonZero(turns.map((t) => t.latency.llm_ttft_ms ?? 0));
+    const ttftP = ttftSamples.length > 0 ? round(percentile(ttftSamples, p), 1) : void 0;
+    const optPct = (key) => {
+      const vals = turns.map((t) => t.latency[key]).filter((v) => typeof v === "number" && v > 0);
+      return vals.length > 0 ? round(percentile(vals, p), 1) : void 0;
+    };
+    const llmTotalP = optPct("llm_total_ms");
+    const endpointP = optPct("endpoint_ms");
+    const bargeinP = optPct("bargein_ms");
+    const ttsTotalP = optPct("tts_total_ms");
     return {
-      stt_ms: round(p95(this._turns.map((t) => t.latency.stt_ms)), 1),
-      llm_ms: round(p95(this._turns.map((t) => t.latency.llm_ms)), 1),
-      tts_ms: round(p95(this._turns.map((t) => t.latency.tts_ms)), 1),
-      total_ms: round(p95(this._turns.map((t) => t.latency.total_ms)), 1)
+      stt_ms: round(percentile(nonZero(turns.map((t) => t.latency.stt_ms)), p), 1),
+      llm_ms: round(percentile(nonZero(turns.map((t) => t.latency.llm_ms)), p), 1),
+      ...ttftP !== void 0 ? { llm_ttft_ms: ttftP } : {},
+      ...llmTotalP !== void 0 ? { llm_total_ms: llmTotalP } : {},
+      tts_ms: round(percentile(nonZero(turns.map((t) => t.latency.tts_ms)), p), 1),
+      total_ms: round(percentile(nonZero(turns.map((t) => t.latency.total_ms)), p), 1),
+      ...endpointP !== void 0 ? { endpoint_ms: endpointP } : {},
+      ...bargeinP !== void 0 ? { bargein_ms: bargeinP } : {},
+      ...ttsTotalP !== void 0 ? { tts_total_ms: ttsTotalP } : {}
     };
   }
 };
@@ -2038,40 +3124,335 @@ function pcm16ToMulaw(pcmData) {
   }
   return out;
 }
+var PcmCarry = class {
+  pending = null;
+  /**
+   * Prepend any carried odd byte, return the even-length prefix, and stash
+   * any new trailing odd byte for the next call.
+   *
+   * Returns a zero-length buffer when no complete sample is yet available.
+   */
+  push(chunk) {
+    const combined = this.pending !== null ? Buffer.concat([this.pending, chunk]) : chunk;
+    this.pending = null;
+    const alignedLen = combined.length & ~1;
+    if (alignedLen < combined.length) {
+      this.pending = combined.subarray(alignedLen);
+    }
+    return combined.subarray(0, alignedLen);
+  }
+  /**
+   * Return any pending byte as a 1-byte buffer (rare in practice — only if
+   * the entire stream had an odd byte count), then reset internal state.
+   */
+  flush() {
+    if (this.pending === null) return Buffer.alloc(0);
+    const out = this.pending;
+    this.pending = null;
+    return out;
+  }
+  /** Reset carry state without flushing. */
+  reset() {
+    this.pending = null;
+  }
+};
+var StatefulResampler = class {
+  srcRate;
+  dstRate;
+  // 16k→8k: 5-tap FIR state.
+  // Extended sample buffer carries the 2 history samples that precede the
+  // current chunk AND any "pending" input sample that did not yet generate
+  // output (i.e. the odd sample when the chunk had an odd sample count).
+  // `firPhase` = 0 means the next output is at input position 0 of the
+  // current chunk; 1 means it starts at input position 1 (because the
+  // previous chunk ended on an even-output boundary).
+  firHistory = new Int16Array(2);
+  // [s_{-2}, s_{-1}]
+  firHistoryValid = false;
+  // Pending sample carried from odd-count chunks (not the byte carry —
+  // this is a complete Int16 sample that becomes the first input for the
+  // next call).
+  firPendingSample = null;
+  // 8k→16k: last input sample deferred across chunk boundaries.
+  upsampleLast = 0;
+  upsampleHasHistory = false;
+  // 24k→16k: fractional phase and last input sample across chunks.
+  resample24Last = 0;
+  resample24Phase = 0;
+  resample24HasHistory = false;
+  // Odd-byte alignment carry.
+  carry = new PcmCarry();
+  constructor(opts) {
+    this.srcRate = opts.srcRate;
+    this.dstRate = opts.dstRate;
+    if (opts.channels !== void 0 && opts.channels !== 1) {
+      throw new Error("StatefulResampler: only mono (channels=1) is supported");
+    }
+    const key = `${this.srcRate}->${this.dstRate}`;
+    if (key !== "16000->8000" && key !== "8000->16000" && key !== "24000->16000") {
+      throw new Error(
+        `StatefulResampler: unsupported conversion ${key}. Supported: 16000->8000, 8000->16000, 24000->16000`
+      );
+    }
+  }
+  /**
+   * Process a chunk of PCM16-LE samples.
+   *
+   * Handles odd-byte inputs via an internal carry buffer. Returns an even-byte-
+   * aligned output buffer; may return a zero-length buffer if not enough
+   * aligned input is available yet.
+   */
+  process(pcm) {
+    const aligned = this.carry.push(pcm);
+    if (aligned.length === 0) return Buffer.alloc(0);
+    if (this.srcRate === 16e3 && this.dstRate === 8e3) {
+      return this._downsample16kTo8k(aligned);
+    }
+    if (this.srcRate === 8e3 && this.dstRate === 16e3) {
+      return this._upsample8kTo16k(aligned);
+    }
+    return this._resample24kTo16k(aligned);
+  }
+  /**
+   * Flush internal state and return any remaining output samples.
+   *
+   * For 8k→16k: the deferred last sample is emitted duplicated (matching
+   * the stateless helper's end-of-stream behaviour).
+   * For 16k→8k: any pending odd sample is processed with edge-replication.
+   * Resets all state after flushing.
+   */
+  flush() {
+    this.carry.flush();
+    if (this.srcRate === 16e3 && this.dstRate === 8e3 && this.firPendingSample !== null) {
+      const s = this.firPendingSample;
+      const tmp = Buffer.alloc(4);
+      tmp.writeInt16LE(s, 0);
+      tmp.writeInt16LE(s, 2);
+      const out = this._downsample16kTo8k(tmp);
+      this.firPendingSample = null;
+      return out;
+    }
+    if (this.srcRate === 8e3 && this.dstRate === 16e3 && this.upsampleHasHistory) {
+      const out = Buffer.alloc(4);
+      out.writeInt16LE(this.upsampleLast, 0);
+      out.writeInt16LE(this.upsampleLast, 2);
+      this.upsampleHasHistory = false;
+      this.upsampleLast = 0;
+      return out;
+    }
+    return Buffer.alloc(0);
+  }
+  /** Reset all carried state (e.g. at call boundaries). */
+  reset() {
+    this.firHistory = new Int16Array(2);
+    this.firHistoryValid = false;
+    this.firPendingSample = null;
+    this.upsampleLast = 0;
+    this.upsampleHasHistory = false;
+    this.resample24Last = 0;
+    this.resample24Phase = 0;
+    this.resample24HasHistory = false;
+    this.carry.reset();
+  }
+  // ---------------------------------------------------------------------------
+  // Private: 16 kHz → 8 kHz
+  // ---------------------------------------------------------------------------
+  /**
+   * 2:1 decimation with a 5-tap binomial FIR anti-alias filter.
+   *
+   * FIR coefficients: [1, 4, 6, 4, 1] / 16 (cutoff ~Fs/4 = 4 kHz).
+   *
+   * Cross-chunk state:
+   * - `firHistory[0]` = s_{-2}, `firHistory[1]` = s_{-1} relative to the
+   *   virtual stream (seeded to first-sample on the very first call).
+   * - `firPendingSample` = a lone input sample carried from a chunk whose
+   *   sample count was odd; it will become the first input of the next chunk.
+   *
+   * Decimation: outputs are at even positions (0, 2, 4 …) in the virtual
+   * extended stream, so every 2 input samples yield 1 output. An odd-sample-
+   * count chunk leaves 1 sample in `firPendingSample`; the next chunk
+   * prepends it so the output cadence is unbroken.
+   */
+  _downsample16kTo8k(buf) {
+    const newSampleCount = buf.length >> 1;
+    const hasPending = this.firPendingSample !== null;
+    const totalInput = newSampleCount + (hasPending ? 1 : 0);
+    const input = new Int16Array(totalInput);
+    if (hasPending) {
+      input[0] = this.firPendingSample;
+      for (let j = 0; j < newSampleCount; j++) input[j + 1] = buf.readInt16LE(j * 2);
+    } else {
+      for (let j = 0; j < newSampleCount; j++) input[j] = buf.readInt16LE(j * 2);
+    }
+    this.firPendingSample = null;
+    if (totalInput === 0) return Buffer.alloc(0);
+    if (!this.firHistoryValid) {
+      this.firHistory[0] = input[0];
+      this.firHistory[1] = input[0];
+      this.firHistoryValid = true;
+    }
+    const extended = new Int16Array(totalInput + 2);
+    extended[0] = this.firHistory[0];
+    extended[1] = this.firHistory[1];
+    for (let j = 0; j < totalInput; j++) extended[j + 2] = input[j];
+    const outSamples = totalInput >> 1;
+    const out = Buffer.alloc(outSamples * 2);
+    for (let i = 0; i < outSamples; i++) {
+      const c = 2 + i * 2;
+      const sM2 = extended[c - 2];
+      const sM1 = extended[c - 1];
+      const s0 = extended[c];
+      const sP1 = c + 1 < extended.length ? extended[c + 1] : extended[extended.length - 1];
+      const sP2 = c + 2 < extended.length ? extended[c + 2] : extended[extended.length - 1];
+      const filtered = sM2 + 4 * sM1 + 6 * s0 + 4 * sP1 + sP2 + 8 >> 4;
+      out.writeInt16LE(Math.max(-32768, Math.min(32767, filtered)), i * 2);
+    }
+    if (totalInput % 2 === 1) {
+      this.firPendingSample = input[totalInput - 1];
+    }
+    if (totalInput >= 2) {
+      this.firHistory[0] = input[totalInput - 2];
+      this.firHistory[1] = input[totalInput - 1];
+    } else {
+      this.firHistory[0] = this.firHistory[1];
+      this.firHistory[1] = input[0];
+    }
+    return out;
+  }
+  // ---------------------------------------------------------------------------
+  // Private: 8 kHz → 16 kHz
+  // ---------------------------------------------------------------------------
+  /**
+   * 1:2 linear-interpolation upsampler.
+   *
+   * For the first chunk (no history): emits 2*(N-1) samples and defers the
+   * last sample. For subsequent chunks (with history): emits the deferred
+   * sample + its interpolated midpoint THEN 2*(N-1) samples from the new
+   * chunk, deferring the new last sample. Total across K chunks + flush =
+   * 2*total_input_samples (correct output length).
+   *
+   * Call flush() after the final chunk to emit the last deferred sample
+   * pair (self-duplicate at end of stream).
+   */
+  _upsample8kTo16k(buf) {
+    const sampleCount = buf.length >> 1;
+    if (sampleCount === 0) return Buffer.alloc(0);
+    const outArr = [];
+    if (this.upsampleHasHistory) {
+      const next = buf.readInt16LE(0);
+      outArr.push(this.upsampleLast);
+      outArr.push(Math.round((this.upsampleLast + next) / 2));
+    }
+    for (let i = 0; i < sampleCount - 1; i++) {
+      const s0 = buf.readInt16LE(i * 2);
+      const s1 = buf.readInt16LE((i + 1) * 2);
+      outArr.push(s0);
+      outArr.push(Math.round((s0 + s1) / 2));
+    }
+    this.upsampleLast = buf.readInt16LE((sampleCount - 1) * 2);
+    this.upsampleHasHistory = true;
+    const outBuf = Buffer.alloc(outArr.length * 2);
+    for (let j = 0; j < outArr.length; j++) outBuf.writeInt16LE(outArr[j], j * 2);
+    return outBuf;
+  }
+  // ---------------------------------------------------------------------------
+  // Private: 24 kHz → 16 kHz
+  // ---------------------------------------------------------------------------
+  /**
+   * 3:2 linear-interpolation decimator (ratio srcRate/dstRate = 1.5).
+   *
+   * `resample24Phase` tracks the fractional input position of the next output
+   * sample relative to the START of the next chunk. Negative phase means the
+   * next output straddles the previous/current chunk boundary; those are
+   * handled using `resample24Last`.
+   */
+  _resample24kTo16k(buf) {
+    const sampleCount = buf.length >> 1;
+    if (sampleCount === 0) return Buffer.alloc(0);
+    const outArr = [];
+    let phase = this.resample24Phase;
+    while (true) {
+      const idx = Math.floor(phase);
+      if (idx >= sampleCount) break;
+      const frac = phase - idx;
+      let s0;
+      let s1;
+      if (idx < 0) {
+        s0 = this.resample24HasHistory ? this.resample24Last : 0;
+        s1 = buf.readInt16LE(0);
+      } else {
+        s0 = buf.readInt16LE(idx * 2);
+        s1 = idx + 1 < sampleCount ? buf.readInt16LE((idx + 1) * 2) : s0;
+      }
+      const interp = Math.round(s0 + (s1 - s0) * frac);
+      outArr.push(Math.max(-32768, Math.min(32767, interp)));
+      phase += 24e3 / 16e3;
+    }
+    this.resample24Last = buf.readInt16LE((sampleCount - 1) * 2);
+    this.resample24HasHistory = true;
+    this.resample24Phase = phase - sampleCount;
+    const outBuf = Buffer.alloc(outArr.length * 2);
+    for (let j = 0; j < outArr.length; j++) outBuf.writeInt16LE(outArr[j], j * 2);
+    return outBuf;
+  }
+};
+function createResampler16kTo8k() {
+  return new StatefulResampler({ srcRate: 16e3, dstRate: 8e3 });
+}
+function createResampler8kTo16k() {
+  return new StatefulResampler({ srcRate: 8e3, dstRate: 16e3 });
+}
+function createResampler24kTo16k() {
+  return new StatefulResampler({ srcRate: 24e3, dstRate: 16e3 });
+}
+var _warnedResample8kTo16k = false;
+var _warnedResample16kTo8k = false;
+var _warnedResample24kTo16k = false;
 function resample8kTo16k(pcm8k) {
-  if (pcm8k.length === 0) return Buffer.alloc(0);
-  const sampleCount = Math.floor(pcm8k.length / 2);
-  const out = Buffer.alloc(sampleCount * 2 * 2);
-  for (let i = 0; i < sampleCount; i++) {
-    const current = pcm8k.readInt16LE(i * 2);
-    const next = i + 1 < sampleCount ? pcm8k.readInt16LE((i + 1) * 2) : current;
-    const interpolated = Math.round((current + next) / 2);
-    out.writeInt16LE(current, i * 4);
-    out.writeInt16LE(interpolated, i * 4 + 2);
+  if (!_warnedResample8kTo16k) {
+    _warnedResample8kTo16k = true;
+    getLogger().warn(
+      "[patter] resample8kTo16k() is deprecated. Use createResampler8kTo16k() (StatefulResampler) to eliminate chunk-boundary discontinuities."
+    );
   }
-  return out;
+  if (pcm8k.length === 0) return Buffer.alloc(0);
+  const r = createResampler8kTo16k();
+  const main = r.process(pcm8k);
+  const tail = r.flush();
+  return tail.length > 0 ? Buffer.concat([main, tail]) : main;
 }
 function resample16kTo8k(pcm16k) {
-  if (pcm16k.length === 0) return Buffer.alloc(0);
-  const sampleCount = Math.floor(pcm16k.length / 2);
-  const outSamples = Math.floor(sampleCount / 2);
-  const out = Buffer.alloc(outSamples * 2);
-  for (let i = 0; i < outSamples; i++) {
-    const sample = pcm16k.readInt16LE(i * 2 * 2);
-    out.writeInt16LE(sample, i * 2);
+  if (!_warnedResample16kTo8k) {
+    _warnedResample16kTo8k = true;
+    getLogger().warn(
+      "[patter] resample16kTo8k() is deprecated. Use createResampler16kTo8k() (StatefulResampler) to eliminate chunk-boundary discontinuities."
+    );
   }
-  return out;
+  if (pcm16k.length === 0) return Buffer.alloc(0);
+  const r = createResampler16kTo8k();
+  const out = r.process(pcm16k);
+  const tail = r.flush();
+  return tail.length > 0 ? Buffer.concat([out, tail]) : out;
 }
 function resample24kTo16k(pcm24k) {
+  if (!_warnedResample24kTo16k) {
+    _warnedResample24kTo16k = true;
+    getLogger().warn(
+      "[patter] resample24kTo16k() is deprecated. Use createResampler24kTo16k() (StatefulResampler) or OpenAITTS.resampleStreaming for anti-aliased resampling."
+    );
+  }
   if (pcm24k.length === 0) return Buffer.alloc(0);
   const sampleCount = Math.floor(pcm24k.length / 2);
   const outSamples = Math.floor(sampleCount * 2 / 3);
   const out = Buffer.alloc(outSamples * 2);
-  let outIdx = 0;
-  for (let i = 0; i < sampleCount && outIdx < outSamples; i++) {
-    if (i % 3 === 2) continue;
-    out.writeInt16LE(pcm24k.readInt16LE(i * 2), outIdx * 2);
-    outIdx++;
+  for (let i = 0; i < outSamples; i++) {
+    const pos = i * 1.5;
+    const idx = Math.floor(pos);
+    const frac = pos - idx;
+    const s0 = pcm24k.readInt16LE(idx * 2);
+    const s1 = idx + 1 < sampleCount ? pcm24k.readInt16LE((idx + 1) * 2) : s0;
+    const interp = Math.round(s0 + (s1 - s0) * frac);
+    out.writeInt16LE(Math.max(-32768, Math.min(32767, interp)), i * 2);
   }
   return out;
 }
@@ -2133,6 +3514,8 @@ async function executeToolWebhook(webhookUrl, toolName, parsedArgs, context, lab
 
 // src/sentence-chunker.ts
 var DEFAULT_MIN_SENTENCE_LEN = 20;
+var DEFAULT_MIN_WORDS_FOR_SHORT_FLUSH = 2;
+var SENTENCE_TERMINATORS = ".!?\u3002\uFF01\uFF1F";
 function splitSentences(text, minSentenceLen = DEFAULT_MIN_SENTENCE_LEN) {
   const alphabets = "([A-Za-z])";
   const prefixes = "(Mr|St|Mrs|Ms|Dr)[.]";
@@ -2191,14 +3574,29 @@ function splitSentences(text, minSentenceLen = DEFAULT_MIN_SENTENCE_LEN) {
 var SentenceChunker = class {
   buffer = "";
   minSentenceLen;
+  minWordsForShortFlush;
   constructor(options) {
     this.minSentenceLen = options?.minSentenceLen ?? DEFAULT_MIN_SENTENCE_LEN;
+    this.minWordsForShortFlush = options?.minWordsForShortFlush ?? DEFAULT_MIN_WORDS_FOR_SHORT_FLUSH;
   }
-  /** Feed a token. Returns zero or more complete sentences. */
+  /**
+   * Feed a token. Returns zero or more complete sentences.
+   *
+   * Two emission paths:
+   * - **Standard path** — when the buffer is at least `minSentenceLen`
+   *   characters long and the regex tokenizer reports more than one
+   *   sentence, all but the last (potentially incomplete) are emitted.
+   * - **Short-flush path** — when the buffer is shorter than `minSentenceLen`
+   *   but ends with a sentence terminator AND has at least
+   *   `minWordsForShortFlush` whitespace-separated words, emit it
+   *   immediately. This drops TTS TTFB on short greetings like `"Hi there!"`
+   *   while keeping single-word utterances (`"Sì."`) buffered until
+   *   `flush()`.
+   */
   push(token) {
     this.buffer += token;
     if (this.buffer.length < this.minSentenceLen) {
-      return [];
+      return this.maybeShortFlush();
     }
     const sentences = splitSentences(this.buffer, this.minSentenceLen);
     if (sentences.length <= 1) {
@@ -2212,6 +3610,41 @@ var SentenceChunker = class {
     this.buffer = sentences[sentences.length - 1]?.[0] ?? "";
     return result;
   }
+  /**
+   * Emit the buffer when it's a short, complete single-sentence utterance.
+   *
+   * A buffer qualifies when **all** of these hold:
+   * 1. Last non-whitespace char is a sentence terminator.
+   * 2. Word count is at least `minWordsForShortFlush` (default 2 — keeps
+   *    single-word "Sì." / "Yes." buffered until `flush()`).
+   * 3. The buffer contains exactly one terminator (the trailing one).
+   *    Multiple terminators mean we may be mid-stream of a longer merged
+   *    utterance like `"Hey! Hi! Hello! This is a sentence."` — let the
+   *    standard path keep merging.
+   * 4. The char immediately before the terminator is NOT a digit (avoids
+   *    decimal mid-stream like `"f(x) = x * 2."` flushing before `54`).
+   * 5. The char immediately before the terminator is NOT an uppercase
+   *    ASCII letter (avoids acronym patterns like `"U.S."` / `"U."`).
+   */
+  maybeShortFlush() {
+    const stripped = this.buffer.replace(/\s+$/, "");
+    if (!stripped) return [];
+    const last = stripped[stripped.length - 1];
+    if (!SENTENCE_TERMINATORS.includes(last)) return [];
+    let terminatorCount = 0;
+    for (const c of stripped) {
+      if (SENTENCE_TERMINATORS.includes(c)) terminatorCount++;
+    }
+    if (terminatorCount !== 1) return [];
+    const wordCount = stripped.split(/\s+/).filter((w) => w.length > 0).length;
+    if (wordCount < this.minWordsForShortFlush) return [];
+    if (stripped.length >= 2) {
+      const prev = stripped[stripped.length - 2];
+      if (/\d/.test(prev) || /[A-Z]/.test(prev)) return [];
+    }
+    this.buffer = "";
+    return [stripped];
+  }
   /** Flush remaining buffer as final sentence(s). Call at end of stream. */
   flush() {
     const remaining = this.buffer.trim();
@@ -2258,6 +3691,44 @@ var PipelineHookExecutor = class {
       return transcript;
     }
   }
+  /**
+   * Run beforeLlm hook. Returns a possibly-modified messages list.
+   * Returning ``null`` from the hook means "keep the original" — the LLM
+   * call is too important to be silently vetoed.
+   * Fail-open: on exception, the original messages pass through.
+   */
+  async runBeforeLlm(messages, ctx) {
+    if (!this.hooks?.beforeLlm) return messages;
+    try {
+      const result = await this.hooks.beforeLlm(messages, ctx);
+      return result ?? messages;
+    } catch (e) {
+      getLogger().error("Pipeline hook beforeLlm threw:", e);
+      return messages;
+    }
+  }
+  /**
+   * Run afterLlm hook. Returns a possibly-modified assistant text.
+   * Returning ``null`` from the hook means "keep the original".
+   * Fail-open: on exception, the original text passes through.
+   */
+  async runAfterLlm(text, ctx) {
+    if (!this.hooks?.afterLlm) return text;
+    try {
+      const result = await this.hooks.afterLlm(text, ctx);
+      return result ?? text;
+    } catch (e) {
+      getLogger().error("Pipeline hook afterLlm threw:", e);
+      return text;
+    }
+  }
+  /**
+   * Whether ``afterLlm`` is configured. Used by the LLM loop to decide
+   * whether to buffer streaming tokens before yielding them.
+   */
+  hasAfterLlm() {
+    return Boolean(this.hooks?.afterLlm);
+  }
   /**
    * Run beforeSynthesize hook. Returns null if hook vetoes TTS for this sentence.
    * If no hook is defined, returns the text unchanged.
@@ -2286,6 +3757,185 @@ var PipelineHookExecutor = class {
   }
 };
 
+// src/observability/event-bus.ts
+var EventBus = class {
+  listeners = /* @__PURE__ */ new Map();
+  /**
+   * Subscribe to an event type.  Returns an unsubscribe function.
+   */
+  on(event, cb) {
+    let set = this.listeners.get(event);
+    if (!set) {
+      set = /* @__PURE__ */ new Set();
+      this.listeners.set(event, set);
+    }
+    set.add(cb);
+    return () => set.delete(cb);
+  }
+  /**
+   * Emit an event synchronously.  Async listeners are fire-and-forget with
+   * rejection logging so a badly-behaved observer never stalls the call path.
+   */
+  emit(event, payload) {
+    const set = this.listeners.get(event);
+    if (!set) return;
+    for (const cb of [...set]) {
+      try {
+        const res = cb(payload);
+        if (res && typeof res.catch === "function") {
+          res.catch(
+            (e) => getLogger().warn(`[EventBus] listener for "${event}" rejected:`, e)
+          );
+        }
+      } catch (e) {
+        getLogger().warn(`[EventBus] listener for "${event}" threw:`, e);
+      }
+    }
+  }
+};
+
+// src/observability/tracing.ts
+var ENV_FLAG = "PATTER_OTEL_ENABLED";
+var SERVICE_NAME = "patter";
+var SPAN_CALL = "getpatter.call";
+var SPAN_STT = "getpatter.stt";
+var SPAN_LLM = "getpatter.llm";
+var SPAN_TTS = "getpatter.tts";
+var SPAN_TOOL = "getpatter.tool";
+var SPAN_ENDPOINT = "getpatter.endpoint";
+var SPAN_BARGEIN = "getpatter.bargein";
+var otel = null;
+var initialized = false;
+var tracerAvailable = false;
+var provider = null;
+function tryLoadOtel() {
+  if (otel !== null) return otel;
+  try {
+    const mod = __require("@opentelemetry/api");
+    otel = mod;
+    return mod;
+  } catch {
+    return null;
+  }
+}
+function trySetupSdk(options, api) {
+  try {
+    const sdkTraceNode = __require("@opentelemetry/sdk-trace-node");
+    const sdkTraceBase = __require("@opentelemetry/sdk-trace-base");
+    const otlpHttp = __require("@opentelemetry/exporter-trace-otlp-http");
+    const serviceName = options.serviceName ?? SERVICE_NAME;
+    const providerInstance = new sdkTraceNode.NodeTracerProvider({
+      resource: {
+        attributes: {
+          "service.name": serviceName,
+          ...options.resourceAttributes ?? {}
+        }
+      }
+    });
+    const endpoint = options.otlpEndpoint ?? process.env.OTEL_EXPORTER_OTLP_ENDPOINT ?? void 0;
+    const exporter = new otlpHttp.OTLPTraceExporter(
+      endpoint ? { url: `${endpoint.replace(/\/$/, "")}/v1/traces` } : void 0
+    );
+    const processor = new sdkTraceBase.BatchSpanProcessor(exporter);
+    providerInstance.addSpanProcessor?.(processor);
+    providerInstance.register?.();
+    try {
+      api.trace.setGlobalTracerProvider?.(providerInstance);
+    } catch {
+    }
+    return providerInstance;
+  } catch (e) {
+    getLogger().debug(
+      `[observability] OTel SDK wire-up skipped: ${String(e?.message ?? e)}`
+    );
+    return null;
+  }
+}
+function envFlagEnabled() {
+  const raw = (process.env[ENV_FLAG] ?? "").toLowerCase();
+  return raw === "1" || raw === "true" || raw === "yes";
+}
+function initTracing(options = {}) {
+  if (initialized) return tracerAvailable;
+  initialized = true;
+  if (!envFlagEnabled()) {
+    tracerAvailable = false;
+    return false;
+  }
+  const api = tryLoadOtel();
+  if (!api) {
+    getLogger().warn(
+      `${ENV_FLAG}=1 but @opentelemetry/api is not installed. Install with: npm install @opentelemetry/api @opentelemetry/sdk-trace-node @opentelemetry/sdk-trace-base @opentelemetry/exporter-trace-otlp-http`
+    );
+    tracerAvailable = false;
+    return false;
+  }
+  provider = trySetupSdk(options, api);
+  tracerAvailable = true;
+  const serviceName = options.serviceName ?? SERVICE_NAME;
+  getLogger().info(
+    `[observability] Patter OTel tracing enabled (service=${serviceName}${provider ? ", exporter=otlp-http" : ", exporter=noop"})`
+  );
+  return true;
+}
+function isTracingEnabled() {
+  return tracerAvailable && envFlagEnabled();
+}
+var NoopSpan = class {
+  setAttribute(_key, _value) {
+  }
+  recordException(_exception) {
+  }
+  end() {
+  }
+};
+var NOOP_SPAN = new NoopSpan();
+var RealSpan = class {
+  span;
+  constructor(span) {
+    this.span = span;
+  }
+  setAttribute(key, value) {
+    try {
+      this.span.setAttribute(key, value);
+    } catch {
+    }
+  }
+  recordException(exception) {
+    try {
+      this.span.recordException(exception);
+    } catch {
+    }
+  }
+  end() {
+    try {
+      this.span.end();
+    } catch {
+    }
+  }
+};
+function startSpan(name, attrs) {
+  if (!isTracingEnabled() || !otel) return NOOP_SPAN;
+  try {
+    const tracer = otel.trace.getTracer(SERVICE_NAME);
+    const rawSpan = tracer.startSpan(name, attrs ? { attributes: attrs } : void 0);
+    return new RealSpan(rawSpan);
+  } catch {
+    return NOOP_SPAN;
+  }
+}
+async function withSpan(name, attrs, fn) {
+  const span = startSpan(name, attrs);
+  try {
+    return await fn(span);
+  } catch (exc) {
+    span.recordException(exc);
+    throw exc;
+  } finally {
+    span.end();
+  }
+}
+
 // src/stream-handler.ts
 function checkGuardrails(text, guardrails) {
   if (!guardrails) return null;
@@ -2305,9 +3955,33 @@ function sanitizeLogValue(v, maxLen = 200) {
   const cleaned = v.replace(/[\x00-\x1f\x7f]/g, "");
   return cleaned.length > maxLen ? cleaned.slice(0, maxLen) + "..." : cleaned;
 }
+function maskPhoneNumber(number) {
+  if (!number) return "***";
+  const text = String(number);
+  if (text.length <= 4) return "***";
+  return `***${text.slice(-4)}`;
+}
 function isValidE164(number) {
   return /^\+[1-9]\d{6,14}$/.test(number);
 }
+var HALLUCINATIONS = /* @__PURE__ */ new Set([
+  "you",
+  "thank you",
+  "thanks",
+  "yeah",
+  "yes",
+  "no",
+  "okay",
+  "ok",
+  "uh",
+  "um",
+  "mmm",
+  "hmm",
+  ".",
+  "bye",
+  "right",
+  "cool"
+]);
 var StreamHandler = class {
   deps;
   ws;
@@ -2320,6 +3994,58 @@ var StreamHandler = class {
   stt = null;
   tts = null;
   isSpeaking = false;
+  /** Set to true after a VAD error to suppress log spam for the rest of the call. */
+  vadDisabled = false;
+  /**
+   * Monotonic counter incremented on every TTS-start. The grace timer
+   * scheduled by ``endSpeakingWithGrace`` only flips ``isSpeaking=false``
+   * if the counter still matches its capture — a new turn that started in
+   * the meantime invalidates the obsolete timer instead of clobbering its
+   * own ``isSpeaking=true``.
+   */
+  speakingGeneration = 0;
+  /** Handle for the pending grace-period timer, so it can be cleared on cleanup. */
+  graceTimer = null;
+  /** Mark the start of a TTS span. Use instead of setting isSpeaking directly. */
+  beginSpeaking() {
+    this.speakingGeneration++;
+    this.isSpeaking = true;
+  }
+  /**
+   * Atomically end speaking AND invalidate any pending grace timer.
+   * Use instead of ``this.isSpeaking = false`` at barge-in sites.
+   */
+  cancelSpeaking() {
+    this.speakingGeneration++;
+    this.isSpeaking = false;
+  }
+  /** Cancel and clear the pending grace timer, if any. */
+  clearGraceTimer() {
+    if (this.graceTimer !== null) {
+      clearTimeout(this.graceTimer);
+      this.graceTimer = null;
+    }
+  }
+  /**
+   * Mark the agent as no longer producing TTS, honoring a grace period that
+   * approximates the carrier's playback buffer. The user may still hear the
+   * agent for ~1 s after we finish pushing audio (Twilio buffers ~1500 ms);
+   * keeping isSpeaking=true through that window keeps the VAD-driven
+   * barge-in armed during the audible tail. Tunable via env.
+   */
+  endSpeakingWithGrace() {
+    const grace = Number(process.env.PATTER_TTS_TAIL_GRACE_MS ?? 1500);
+    if (grace > 0) {
+      const gen = this.speakingGeneration;
+      this.clearGraceTimer();
+      this.graceTimer = setTimeout(() => {
+        this.graceTimer = null;
+        if (this.speakingGeneration === gen) this.isSpeaking = false;
+      }, grace);
+    } else {
+      this.isSpeaking = false;
+    }
+  }
   llmLoop = null;
   chunkCount = 0;
   callEndFired = false;
@@ -2329,30 +4055,155 @@ var StreamHandler = class {
   maxDurationTimer = null;
   transcriptProcessing = false;
   transcriptQueue = [];
-  // BUG #22 throttle state — mirror Python impl.
+  // Throttle state for back-to-back STT finals — see ``commitTranscript``.
   lastCommitText = "";
   lastCommitAt = 0;
+  // PCM16 byte-alignment carry for TTS streaming (pipeline mode).
+  // HTTP streams from ElevenLabs / OpenAI / Cartesia can yield chunks of any
+  // size, including odd byte counts. Silently dropping the trailing odd byte
+  // misaligns every subsequent int16 sample in the stream (hi/lo bytes get
+  // swapped), producing a voice drowned in loud hiss. We buffer the odd byte
+  // across chunks so resample/mulaw encoding always sees aligned int16 frames.
+  ttsByteCarry = null;
+  // Per-session stateful resamplers eliminate chunk-boundary discontinuities.
+  // Created lazily on first use; reset() on call end.
+  inboundResampler = createResampler8kTo16k();
+  outboundResampler = createResampler16kTo8k();
   history;
   metricsAcc;
+  _eventBus;
   constructor(deps, ws, caller, callee) {
     this.deps = deps;
     this.ws = ws;
     this.caller = caller;
     this.callee = callee;
     this.history = createHistoryManager(200);
-    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 sttKey = deps.agent.stt?.constructor?.providerKey;
+    const sttProviderName = deps.agent.stt ? sttKey ?? deps.agent.stt.constructor?.name ?? "custom" : void 0;
+    const ttsKey = deps.agent.tts?.constructor?.providerKey;
+    const ttsProviderName = deps.agent.tts ? ttsKey ?? deps.agent.tts.constructor?.name ?? "custom" : void 0;
     const providerMode = deps.agent.provider ?? "openai_realtime";
+    const llmKey = deps.agent.llm?.constructor?.providerKey;
+    let llmProviderName;
+    if (deps.agent.llm) {
+      if (llmKey) {
+        llmProviderName = llmKey;
+      } else {
+        const stripped = (deps.agent.llm.constructor?.name ?? "custom").replace(/LLMProvider$/i, "").replace(/LLM$/i, "").replace(/Provider$/i, "").toLowerCase();
+        llmProviderName = stripped || "custom";
+      }
+    } else {
+      llmProviderName = providerMode === "openai_realtime" ? "openai_realtime" : "openai";
+    }
+    this._eventBus = new EventBus();
     this.metricsAcc = new CallMetricsAccumulator({
       callId: "",
       providerMode,
       telephonyProvider: deps.bridge.telephonyProvider,
       sttProvider: sttProviderName,
       ttsProvider: ttsProviderName,
-      pricing: deps.pricing
+      llmProvider: llmProviderName,
+      pricing: deps.pricing,
+      eventBus: this._eventBus,
+      reportOnlyInitialTtfb: deps.reportOnlyInitialTtfb ?? false
     });
     getLogger().debug(`WebSocket connection opened (${deps.bridge.label})`);
   }
+  /**
+   * Record a completed turn in the dashboard store and fire the user-supplied
+   * ``onMetrics`` callback. Centralises the 4 emit sites (firstMessage, pipeline
+   * streaming/regular LLM, WebSocket remote, Realtime response_done) so the
+   * payload shape lives in one place.
+   */
+  async emitTurnMetrics(turn) {
+    if (turn == null) return;
+    this.deps.metricsStore.recordTurn({ call_id: this.callId, turn });
+    if (!this.deps.onMetrics) return;
+    const turnMetrics = turn;
+    const llm_ttft_ms = turnMetrics?.latency?.llm_ttft_ms;
+    await this.deps.onMetrics({
+      call_id: this.callId,
+      turn,
+      ...llm_ttft_ms !== void 0 ? { llm_ttft_ms } : {},
+      cost_so_far: this.metricsAcc.getCostSoFar()
+    });
+  }
+  /** Reset the TTS odd-byte carry — call at every TTS stream entry/exit. */
+  resetTtsCarry() {
+    this.ttsByteCarry = null;
+  }
+  /**
+   * Flush both stateful resamplers and any TTS byte carry on call close.
+   * Emits tail bytes through the telephony bridge so the last ~20 ms of audio
+   * is not silently clipped on hangup. No-op if the WebSocket is already gone.
+   */
+  flushResamplers() {
+    try {
+      const inTail = this.inboundResampler.flush();
+      if (inTail.length > 0 && this.stt) {
+        this.stt.sendAudio(inTail);
+      }
+    } catch {
+    }
+    try {
+      const outTail = this.outboundResampler.flush();
+      if (outTail.length > 0 && this.ws.readyState === this.ws.OPEN) {
+        const mulaw = pcm16ToMulaw(outTail);
+        this.deps.bridge.sendAudio(this.ws, mulaw.toString("base64"), this.streamSid);
+      }
+    } catch {
+    }
+    this.ttsByteCarry = null;
+  }
+  /**
+   * Start call recording when configured. Currently Twilio-only — bridges may
+   * expose ``startRecording`` for parity when we add other carriers.
+   */
+  async startRecordingIfRequested(callId) {
+    const { recording, config } = this.deps;
+    if (!recording || !config.twilioSid || !config.twilioToken || !callId) return;
+    if (!validateTwilioSid(callId)) {
+      getLogger().warn(`Recording skipped: invalid Twilio CallSid format ${JSON.stringify(callId)}`);
+      return;
+    }
+    try {
+      const recUrl = `https://api.twilio.com/2010-04-01/Accounts/${config.twilioSid}/Calls/${callId}/Recordings.json`;
+      const recResp = await fetch(recUrl, {
+        method: "POST",
+        headers: {
+          "Authorization": `Basic ${Buffer.from(`${config.twilioSid}:${config.twilioToken}`).toString("base64")}`
+        }
+      });
+      if (recResp.ok) {
+        getLogger().debug(`Recording started for ${callId}`);
+      } else {
+        getLogger().warn(`could not start recording: ${await recResp.text()}`);
+      }
+    } catch (e) {
+      getLogger().warn(`could not start recording: ${String(e)}`);
+    }
+  }
+  // ---------------------------------------------------------------------------
+  // Public: observer API
+  // ---------------------------------------------------------------------------
+  /**
+   * Subscribe to a Patter event on the per-call EventBus.
+   *
+   * The most common use-case is 'metrics_collected' — fired after every
+   * completed turn with the TurnMetrics payload.
+   *
+   * Returns an unsubscribe function; call it to stop receiving events.
+   *
+   * @example
+   * const off = handler.addObserver((payload) => {
+   *   console.log('turn metrics:', payload);
+   * });
+   * // later:
+   * off();
+   */
+  addObserver(cb, event = "metrics_collected") {
+    return this._eventBus.on(event, cb);
+  }
   // ---------------------------------------------------------------------------
   // Public: called by the provider-specific parsers in server.ts
   // ---------------------------------------------------------------------------
@@ -2377,8 +4228,7 @@ var StreamHandler = class {
     this.deps.metricsStore.recordCallStart({
       call_id: callId,
       caller: this.caller,
-      callee: this.callee,
-      direction: "inbound"
+      callee: this.callee
     });
     const MAX_CALL_DURATION_MS = 60 * 60 * 1e3;
     this.maxDurationTimer = setTimeout(async () => {
@@ -2389,52 +4239,32 @@ var StreamHandler = class {
       }
     }, MAX_CALL_DURATION_MS);
     try {
-      const { notifyDashboard } = await import("./persistence-CYIGNHSU.mjs");
+      const { notifyDashboard } = await import("./persistence-LQBYQPQQ.mjs");
       notifyDashboard({
         call_id: callId,
         caller: this.caller,
-        callee: this.callee,
-        direction: "inbound"
+        callee: this.callee
       });
     } catch {
     }
     if (this.deps.onCallStart) {
+      const direction = this.deps.metricsStore.getActive(callId)?.direction ?? "inbound";
       await this.deps.onCallStart({
         call_id: callId,
         caller: this.caller,
         callee: this.callee,
-        direction: "inbound",
+        direction,
+        telephony_provider: this.deps.bridge.telephonyProvider,
         ...Object.keys(customParams).length > 0 ? { custom_params: customParams } : {}
       });
     }
-    if (this.deps.recording && this.deps.config.twilioSid && this.deps.config.twilioToken && callId) {
-      if (!validateTwilioSid(callId)) {
-        getLogger().warn(`Recording skipped: invalid Twilio CallSid format ${JSON.stringify(callId)}`);
-      } else {
-        try {
-          const recUrl = `https://api.twilio.com/2010-04-01/Accounts/${this.deps.config.twilioSid}/Calls/${callId}/Recordings.json`;
-          const recResp = await fetch(recUrl, {
-            method: "POST",
-            headers: {
-              "Authorization": `Basic ${Buffer.from(`${this.deps.config.twilioSid}:${this.deps.config.twilioToken}`).toString("base64")}`
-            }
-          });
-          if (recResp.ok) {
-            getLogger().debug(`Recording started for ${callId}`);
-          } else {
-            getLogger().warn(`could not start recording: ${await recResp.text()}`);
-          }
-        } catch (e) {
-          getLogger().warn(`could not start recording: ${String(e)}`);
-        }
-      }
-    }
+    await this.startRecordingIfRequested(callId);
     const agentVars = this.deps.sanitizeVariables(this.deps.agent.variables ?? {});
     const safeCustomParams = this.deps.sanitizeVariables(customParams);
     const allVars = { ...agentVars, ...safeCustomParams };
     const resolvedPrompt = Object.keys(allVars).length > 0 ? this.deps.resolveVariables(this.deps.agent.systemPrompt, allVars) : this.deps.agent.systemPrompt;
-    const provider = this.deps.agent.provider ?? "openai_realtime";
-    if (provider === "pipeline") {
+    const provider2 = this.deps.agent.provider ?? "openai_realtime";
+    if (provider2 === "pipeline") {
       await this.initPipeline(resolvedPrompt);
     } else {
       await this.initRealtimeAdapter(resolvedPrompt);
@@ -2446,13 +4276,56 @@ var StreamHandler = class {
   }
   /** Handle an incoming audio chunk (already decoded from base64). */
   async handleAudio(audioBuffer) {
-    const provider = this.deps.agent.provider ?? "openai_realtime";
-    if (provider === "pipeline" && this.stt) {
-      if (this.isSpeaking && (this.deps.agent.bargeInThresholdMs ?? 300) === 0) {
-        return;
-      }
+    const provider2 = this.deps.agent.provider ?? "openai_realtime";
+    if (provider2 === "pipeline" && this.stt) {
       const pcm8k = mulawToPcm16(audioBuffer);
-      const pcm16k = resample8kTo16k(pcm8k);
+      const pcm16k = this.inboundResampler.process(pcm8k);
+      if (this.deps.agent.vad && !this.vadDisabled) {
+        try {
+          const vadPromise = this.deps.agent.vad.processFrame(pcm16k, 16e3);
+          const timeoutPromise = new Promise((resolve) => setTimeout(() => resolve(null), 25));
+          const evt = await Promise.race([vadPromise, timeoutPromise]);
+          if (evt) {
+            getLogger().info(
+              `[VAD] ${evt.type}  agentSpeaking=${this.isSpeaking}`
+            );
+          }
+          if (evt?.type === "speech_start") {
+            if (this.isSpeaking) {
+              getLogger().info("[VAD] speech_start during TTS \u2192 BARGE-IN");
+              this.metricsAcc.recordOverlapStart();
+              this.metricsAcc.recordBargeinDetected();
+              const bargeinSpan = startSpan(SPAN_BARGEIN, { "patter.call.id": this.callId });
+              try {
+                this.cancelSpeaking();
+                try {
+                  this.deps.bridge.sendClear(this.ws, this.streamSid);
+                } catch (err) {
+                  getLogger().debug(`sendClear during VAD barge-in failed: ${String(err)}`);
+                }
+                this.metricsAcc.recordTtsStopped();
+                this.metricsAcc.recordTurnInterrupted();
+                this.metricsAcc.recordOverlapEnd(true);
+              } finally {
+                try {
+                  bargeinSpan.end();
+                } catch {
+                }
+              }
+            }
+            this.metricsAcc.startTurnIfIdle();
+          } else if (evt?.type === "speech_end") {
+            this.metricsAcc.recordVadStop();
+          }
+        } catch (err) {
+          this.vadDisabled = true;
+          getLogger().warn(`VAD processFrame failed \u2014 disabling VAD for this call: ${String(err)}`);
+        }
+      }
+      if (this.isSpeaking) {
+        if (this.deps.agent.vad) return;
+        if ((this.deps.agent.bargeInThresholdMs ?? 300) === 0) return;
+      }
       const hooks = this.deps.agent.hooks;
       if (hooks) {
         const hookExecutor = new PipelineHookExecutor(hooks);
@@ -2460,13 +4333,15 @@ var StreamHandler = class {
         const processed = await hookExecutor.runBeforeSendToStt(pcm16k, hookCtx);
         if (processed === null) return;
         this.stt.sendAudio(processed);
+        this.metricsAcc.addSttAudioBytes(processed.length);
       } else {
         this.stt.sendAudio(pcm16k);
+        this.metricsAcc.addSttAudioBytes(pcm16k.length);
       }
     } else if (this.adapter) {
-      if (this.adapter instanceof ElevenLabsConvAIAdapter && this.deps.bridge.telephonyProvider === "twilio") {
+      if (this.adapter instanceof ElevenLabsConvAIAdapter && this.deps.bridge.telephonyProvider === "twilio" && this.adapter.inputAudioFormat !== "ulaw_8000") {
         const pcm8k = mulawToPcm16(audioBuffer);
-        const pcm16k = resample8kTo16k(pcm8k);
+        const pcm16k = this.inboundResampler.process(pcm8k);
         this.adapter.sendAudio(pcm16k);
       } else {
         this.adapter.sendAudio(audioBuffer);
@@ -2483,8 +4358,28 @@ var StreamHandler = class {
       await this.deps.onTranscript({ role: "user", text: `[DTMF: ${digit}]`, call_id: this.callId });
     }
   }
+  /**
+   * Last mark name Twilio has confirmed playback of. Mirrors the Python
+   * ``TwilioAudioSender.last_confirmed_mark`` field — barge-in heuristics
+   * compare this against the latest sent mark to decide whether the agent's
+   * audio has actually reached the caller yet.
+   */
+  lastConfirmedMark = "";
+  /**
+   * Handle a Twilio ``mark`` event acknowledging that a previously sent
+   * audio chunk has been played out. Mirrors Python's
+   * ``twilio_handler.py``: ``audio_sender.on_mark_confirmed(mark_name)`` +
+   * ``handler.on_mark(mark_name)``.
+   */
+  async onMark(markName) {
+    if (markName) {
+      this.lastConfirmedMark = markName;
+    }
+  }
   /** Handle call stop / stream end. */
   async handleStop() {
+    this.clearGraceTimer();
+    this.flushResamplers();
     await this.closeSttOnce();
     try {
       this.adapter?.close();
@@ -2494,6 +4389,8 @@ var StreamHandler = class {
   }
   /** Handle WebSocket close event. */
   async handleWsClose() {
+    this.clearGraceTimer();
+    this.flushResamplers();
     await this.closeSttOnce();
     try {
       this.adapter?.close();
@@ -2519,15 +4416,33 @@ var StreamHandler = class {
   // ---------------------------------------------------------------------------
   /**
    * Encode a PCM 16kHz audio chunk for the telephony provider.
-   * Twilio requires mulaw 8kHz; Telnyx accepts PCM 16kHz natively.
+   *
+   * Both Twilio and Telnyx negotiate PCMU (mulaw) 8 kHz on the bidirectional
+   * media stream — Twilio always, and Telnyx because ``streaming_start``
+   * (server.ts) requests ``stream_bidirectional_codec=PCMU`` at 8 kHz. So
+   * the wire format for both providers is mulaw 8 kHz; we resample 16 kHz
+   * PCM16 → 8 kHz then encode to mulaw. Mirrors the Python pipeline path
+   * (sdk-py/getpatter/handlers/telnyx_handler.py::TelnyxAudioSender).
+   *
+   * Maintains a 1-byte carry across calls so unaligned HTTP chunks from
+   * streaming TTS providers never byte-swap the PCM16 samples downstream.
    */
   encodePipelineAudio(pcm16k) {
-    if (this.deps.bridge.telephonyProvider === "twilio") {
-      const pcm8k = resample16kTo8k(pcm16k);
-      const mulaw = pcm16ToMulaw(pcm8k);
-      return mulaw.toString("base64");
-    }
-    return pcm16k.toString("base64");
+    const aligned = this.alignPcm16(pcm16k);
+    if (aligned.length === 0) return "";
+    const pcm8k = this.outboundResampler.process(aligned);
+    const mulaw = pcm16ToMulaw(pcm8k);
+    return mulaw.toString("base64");
+  }
+  /**
+   * Prepend any carry byte from the previous chunk, return the even-length
+   * portion, and stash the final odd byte (if any) for the next call.
+   */
+  alignPcm16(chunk) {
+    const combined = this.ttsByteCarry ? Buffer.concat([this.ttsByteCarry, chunk]) : chunk;
+    const alignedLen = combined.length & ~1;
+    this.ttsByteCarry = alignedLen < combined.length ? combined.subarray(alignedLen) : null;
+    return combined.subarray(0, alignedLen);
   }
   // ---------------------------------------------------------------------------
   // Private: Pipeline mode
@@ -2556,6 +4471,7 @@ var StreamHandler = class {
     if (this.deps.agent.firstMessage && !this.deps.onMessage && this.tts) {
       this.metricsAcc.startTurn();
       let firstChunkSent = false;
+      this.resetTtsCarry();
       try {
         for await (const chunk of this.tts.synthesizeStream(this.deps.agent.firstMessage)) {
           if (!firstChunkSent) {
@@ -2567,13 +4483,11 @@ var StreamHandler = class {
         }
       } catch (e) {
         getLogger().error(`First message TTS error (${label}):`, e);
+      } finally {
+        this.resetTtsCarry();
       }
       if (firstChunkSent) {
-        const turn = this.metricsAcc.recordTurnComplete(this.deps.agent.firstMessage);
-        if (turn) {
-          this.deps.metricsStore.recordTurn({ call_id: this.callId, turn });
-          if (this.deps.onMetrics) await this.deps.onMetrics({ call_id: this.callId, turn });
-        }
+        await this.emitTurnMetrics(this.metricsAcc.recordTurnComplete(this.deps.agent.firstMessage));
         this.history.push({ role: "assistant", text: this.deps.agent.firstMessage, timestamp: Date.now() });
       }
     }
@@ -2583,15 +4497,17 @@ var StreamHandler = class {
           "Cannot pass both agent({ llm }) and serve({ onMessage }). Pick one \u2014 `llm` for built-in LLMs, `onMessage` for custom logic."
         );
       }
+      const providerModel = this.deps.agent.llm?.model ?? "";
       this.llmLoop = new LLMLoop(
         "",
         // apiKey unused when llmProvider is supplied
-        "",
-        // model unused when llmProvider is supplied
+        providerModel,
+        // propagate so calculateLlmCost can match the price row
         resolvedPrompt,
         this.deps.agent.tools,
         this.deps.agent.llm
       );
+      this.llmLoop.setEventBus(this._eventBus);
       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) {
@@ -2603,6 +4519,7 @@ var StreamHandler = class {
         resolvedPrompt,
         this.deps.agent.tools
       );
+      this.llmLoop.setEventBus(this._eventBus);
       getLogger().debug(`Built-in LLM loop active (pipeline, ${label})`);
     }
     if (this.stt) {
@@ -2632,6 +4549,7 @@ var StreamHandler = class {
     }
     const processedText = await hookExecutor.runBeforeSynthesize(transformed, hookCtx);
     if (processedText === null) return;
+    this.resetTtsCarry();
     try {
       for await (const chunk of this.tts.synthesizeStream(processedText)) {
         if (!this.isSpeaking) break;
@@ -2646,6 +4564,8 @@ var StreamHandler = class {
       }
     } catch (e) {
       getLogger().error(`TTS streaming error (${this.deps.bridge.label}):`, e);
+    } finally {
+      this.resetTtsCarry();
     }
   }
   /** Handle a final transcript from STT in pipeline mode. */
@@ -2663,63 +4583,30 @@ var StreamHandler = class {
     }
   }
   async processTranscript(transcript) {
-    if (transcript.text && this.isSpeaking) {
-      getLogger().debug(
-        `Barge-in: caller spoke over agent (${sanitizeLogValue(transcript.text.slice(0, 40))})`
-      );
-      this.isSpeaking = false;
-      try {
-        this.deps.bridge.sendClear(this.ws, this.streamSid);
-      } catch (err) {
-        getLogger().debug(`sendClear during barge-in failed: ${String(err)}`);
-      }
-      this.metricsAcc.recordTurnInterrupted();
+    let interrupted = this.handleBargeIn(transcript);
+    if (transcript.text) {
+      this.metricsAcc.startTurnIfIdle();
     }
-    if (!transcript.isFinal || !transcript.text) return;
-    const now = Date.now();
-    const normalised = transcript.text.trim().toLowerCase();
-    const stripped = normalised.replace(/[.,!?;: ]+$/, "").trim();
-    const sinceLastMs = now - this.lastCommitAt;
-    const HALLUCINATIONS = /* @__PURE__ */ new Set([
-      "you",
-      "thank you",
-      "thanks",
-      "yeah",
-      "yes",
-      "no",
-      "okay",
-      "ok",
-      "uh",
-      "um",
-      "mmm",
-      "hmm",
-      ".",
-      "bye",
-      "right",
-      "cool"
-    ]);
-    if (HALLUCINATIONS.has(stripped) || stripped === "") {
-      getLogger().debug(`Dropped likely STT hallucination: ${sanitizeLogValue(normalised.slice(0, 40))}`);
-      return;
+    if (transcript.speechFinal) {
+      this.metricsAcc.recordVadStop();
     }
-    if (sinceLastMs < 2e3 && normalised === this.lastCommitText) {
-      getLogger().debug(
-        `Dropped duplicate final transcript (${(sinceLastMs / 1e3).toFixed(1)}s since last): ${sanitizeLogValue(normalised.slice(0, 40))}`
-      );
-      return;
-    }
-    if (sinceLastMs < 500) {
-      getLogger().debug(
-        `Dropped back-to-back final transcript (${(sinceLastMs / 1e3).toFixed(2)}s since last): ${sanitizeLogValue(normalised.slice(0, 40))}`
-      );
-      return;
-    }
-    this.lastCommitText = normalised;
-    this.lastCommitAt = now;
+    if (!transcript.isFinal || !transcript.text) return;
+    if (!this.commitTranscript(transcript.text)) return;
     const label = this.deps.bridge.label;
     getLogger().debug(`User (${label} pipeline): ${sanitizeLogValue(transcript.text)}`);
-    this.metricsAcc.startTurn();
+    this.metricsAcc.startTurnIfIdle();
     this.metricsAcc.recordSttComplete(transcript.text);
+    this.metricsAcc.recordSttFinalTimestamp();
+    const endpointSpan = startSpan(SPAN_ENDPOINT, { "patter.call.id": this.callId });
+    let endpointSpanClosed = false;
+    const closeEndpointSpan = () => {
+      if (endpointSpanClosed) return;
+      endpointSpanClosed = true;
+      try {
+        endpointSpan.end();
+      } catch {
+      }
+    };
     if (this.deps.onTranscript) {
       await this.deps.onTranscript({
         role: "user",
@@ -2734,10 +4621,14 @@ var StreamHandler = class {
     if (filteredTranscript === null) {
       getLogger().debug(`afterTranscribe hook vetoed turn (${label})`);
       this.metricsAcc.recordTurnInterrupted();
+      closeEndpointSpan();
       return;
     }
     this.history.push({ role: "user", text: filteredTranscript, timestamp: Date.now() });
     let responseText = "";
+    this.metricsAcc.recordOnUserTurnCompletedDelay(0);
+    this.metricsAcc.recordTurnCommitted();
+    closeEndpointSpan();
     if (this.deps.onMessage && typeof this.deps.onMessage === "function") {
       try {
         responseText = await this.deps.onMessage({
@@ -2767,104 +4658,203 @@ var StreamHandler = class {
       if (isWebSocketUrl(this.deps.onMessage)) {
         await this.handleWebSocketResponse(msgData);
         return;
-      } else {
-        try {
-          responseText = await this.deps.remoteHandler.callWebhook(this.deps.onMessage, msgData);
-        } catch (e) {
-          getLogger().error(`Webhook remote error (${label}):`, e);
-          return;
-        }
       }
-    } else if (this.llmLoop) {
-      const callCtx = { call_id: this.callId, caller: this.caller, callee: this.callee };
-      const chunker = new SentenceChunker();
-      const allParts = [];
-      const ttsFirstByteSent = { value: false };
-      this.isSpeaking = true;
-      let llmError = false;
       try {
-        try {
-          for await (const token of this.llmLoop.run(filteredTranscript, this.history.entries, callCtx)) {
-            allParts.push(token);
-            const sentences = chunker.push(token);
-            for (const sentence of sentences) {
-              if (!this.isSpeaking) break;
-              const guard = checkGuardrails(sentence, this.deps.agent.guardrails);
-              const sentenceText = guard ? guard.replacement ?? "I'm sorry, I can't respond to that." : sentence;
-              await this.synthesizeSentence(sentenceText, hookExecutor, hookCtx, ttsFirstByteSent);
-            }
-            if (!this.isSpeaking) break;
-          }
-        } catch (e) {
-          llmError = true;
-          chunker.reset();
-          getLogger().error(`LLM loop error (${label}):`, e);
-        }
-        this.metricsAcc.recordLlmComplete();
-        if (!llmError && this.isSpeaking) {
-          for (const sentence of chunker.flush()) {
-            if (!this.isSpeaking) break;
-            const guard = checkGuardrails(sentence, this.deps.agent.guardrails);
-            const sentenceText = guard ? guard.replacement ?? "I'm sorry, I can't respond to that." : sentence;
-            await this.synthesizeSentence(sentenceText, hookExecutor, hookCtx, ttsFirstByteSent);
-          }
-        }
-      } finally {
-        this.isSpeaking = false;
+        responseText = await this.deps.remoteHandler.callWebhook(this.deps.onMessage, msgData);
+      } catch (e) {
+        getLogger().error(`Webhook remote error (${label}):`, e);
+        return;
       }
-      responseText = allParts.join("");
+    } else if (this.llmLoop) {
+      responseText = await this.runPipelineLlm(filteredTranscript, hookExecutor, hookCtx);
     } else {
       return;
     }
     if (!responseText) return;
-    if (!this.llmLoop) {
-      const guard = checkGuardrails(responseText, this.deps.agent.guardrails);
-      if (guard) {
-        getLogger().debug(`Guardrail '${guard.name}' triggered (pipeline)`);
-        responseText = guard.replacement ?? "I'm sorry, I can't respond to that.";
-      }
-      this.metricsAcc.recordLlmComplete();
+    if (this.llmLoop) {
       this.history.push({ role: "assistant", text: responseText, timestamp: Date.now() });
-      const chunker = new SentenceChunker();
-      const sentences = [...chunker.push(responseText), ...chunker.flush()];
-      const ttsFirstByteSent = { value: false };
-      let interrupted = false;
-      this.isSpeaking = true;
+      this.metricsAcc.recordTtsComplete(responseText);
+    } else {
+      interrupted = await this.runRegularLlm(responseText, hookExecutor, hookCtx) || interrupted;
+      responseText = this.history.entries[this.history.entries.length - 1]?.text ?? responseText;
+    }
+    if (!interrupted) {
+      await this.emitTurnMetrics(this.metricsAcc.recordTurnComplete(responseText));
+    }
+  }
+  /**
+   * Barge-in: caller spoke over in-flight TTS. Flip ``isSpeaking`` so the
+   * sentence loop exits on its next check, clear downstream audio buffers,
+   * record the interruption, and return ``true`` so the caller skips the
+   * turn-complete record.
+   */
+  handleBargeIn(transcript) {
+    if (!transcript.text || !this.isSpeaking) return false;
+    getLogger().debug(
+      `Barge-in: caller spoke over agent (${sanitizeLogValue(transcript.text.slice(0, 40))})`
+    );
+    this.metricsAcc.recordOverlapStart();
+    this.metricsAcc.recordBargeinDetected();
+    const bargeinSpan = startSpan(SPAN_BARGEIN, { "patter.call.id": this.callId });
+    try {
+      this.cancelSpeaking();
+      try {
+        this.deps.bridge.sendClear(this.ws, this.streamSid);
+      } catch (err) {
+        getLogger().debug(`sendClear during barge-in failed: ${String(err)}`);
+      }
+      this.metricsAcc.recordTtsStopped();
+      this.metricsAcc.recordTurnInterrupted();
+      this.metricsAcc.recordOverlapEnd(true);
+    } finally {
       try {
-        for (const sentence of sentences) {
-          if (!this.isSpeaking) {
-            interrupted = true;
-            break;
+        bargeinSpan.end();
+      } catch {
+      }
+    }
+    return true;
+  }
+  /**
+   * Dedup + throttle + hallucination filter for final STT transcripts.
+   * Mirrors ``PipelineStreamHandler._stt_loop`` on the Python side.
+   * Returns ``true`` when the transcript should be committed to a turn,
+   * ``false`` when it must be dropped. Drop reasons:
+   *   - text matches common short hallucinations ("you", "thanks", ...)
+   *   - duplicate final within 2 s of previous commit
+   *   - back-to-back finals under 500 ms (too tight to be real utterances)
+   */
+  commitTranscript(text) {
+    const now = Date.now();
+    const normalised = text.trim().toLowerCase();
+    const stripped = normalised.replace(/[.,!?;: ]+$/, "").trim();
+    const sinceLastMs = now - this.lastCommitAt;
+    if (HALLUCINATIONS.has(stripped) || stripped === "") {
+      getLogger().debug(`Dropped likely STT hallucination: ${sanitizeLogValue(normalised.slice(0, 40))}`);
+      return false;
+    }
+    if (sinceLastMs < 2e3 && normalised === this.lastCommitText) {
+      getLogger().debug(
+        `Dropped duplicate final transcript (${(sinceLastMs / 1e3).toFixed(1)}s since last): ${sanitizeLogValue(normalised.slice(0, 40))}`
+      );
+      return false;
+    }
+    if (sinceLastMs < 500) {
+      getLogger().debug(
+        `Dropped back-to-back final transcript (${(sinceLastMs / 1e3).toFixed(2)}s since last): ${sanitizeLogValue(normalised.slice(0, 40))}`
+      );
+      return false;
+    }
+    this.lastCommitText = normalised;
+    this.lastCommitAt = now;
+    return true;
+  }
+  /**
+   * Streaming built-in LLM path with sentence chunking and per-sentence
+   * guardrails/TTS. Returns the concatenated response text.
+   */
+  async runPipelineLlm(filteredTranscript, hookExecutor, hookCtx) {
+    const label = this.deps.bridge.label;
+    const callCtx = { call_id: this.callId, caller: this.caller, callee: this.callee };
+    const chunker = new SentenceChunker();
+    const allParts = [];
+    const ttsFirstByteSent = { value: false };
+    this.beginSpeaking();
+    let llmError = false;
+    const llmSpan = startSpan(SPAN_LLM, { "patter.call.id": this.callId });
+    const guardAndSpeak = async (sentence, isFirst) => {
+      if (isFirst) this.metricsAcc.recordLlmFirstSentenceComplete();
+      const guard = checkGuardrails(sentence, this.deps.agent.guardrails);
+      const sentenceText = guard ? guard.replacement ?? "I'm sorry, I can't respond to that." : sentence;
+      await this.synthesizeSentence(sentenceText, hookExecutor, hookCtx, ttsFirstByteSent);
+    };
+    let firstSentenceEmitted = false;
+    try {
+      try {
+        for await (const token of this.llmLoop.run(
+          filteredTranscript,
+          this.history.entries,
+          callCtx,
+          this.metricsAcc,
+          hookExecutor,
+          hookCtx
+        )) {
+          this.metricsAcc.recordLlmFirstToken();
+          allParts.push(token);
+          for (const sentence of chunker.push(token)) {
+            if (!this.isSpeaking) break;
+            await guardAndSpeak(sentence, !firstSentenceEmitted);
+            firstSentenceEmitted = true;
           }
-          await this.synthesizeSentence(sentence, hookExecutor, hookCtx, ttsFirstByteSent);
+          if (!this.isSpeaking) break;
         }
-      } finally {
-        this.isSpeaking = false;
+      } catch (e) {
+        llmError = true;
+        chunker.reset();
+        getLogger().error(`LLM loop error (${label}):`, e);
+        this.metricsAcc.recordTurnInterrupted();
       }
-      if (!interrupted) {
-        this.metricsAcc.recordTtsComplete(responseText);
+      this.metricsAcc.recordLlmComplete();
+      if (!llmError && this.isSpeaking) {
+        for (const sentence of chunker.flush()) {
+          if (!this.isSpeaking) break;
+          await guardAndSpeak(sentence, !firstSentenceEmitted);
+          firstSentenceEmitted = true;
+        }
       }
-    } else {
-      this.history.push({ role: "assistant", text: responseText, timestamp: Date.now() });
-      this.metricsAcc.recordTtsComplete(responseText);
+    } finally {
+      this.endSpeakingWithGrace();
+      try {
+        llmSpan.end();
+      } catch {
+      }
+    }
+    return allParts.join("");
+  }
+  /**
+   * Non-streaming path (onMessage function / webhook): apply output guardrails,
+   * push to history, sentence-chunk the text, synthesize. Returns ``true`` if
+   * TTS was interrupted mid-flight so the caller can skip turn-complete.
+   */
+  async runRegularLlm(responseText, hookExecutor, hookCtx) {
+    const guard = checkGuardrails(responseText, this.deps.agent.guardrails);
+    let text = responseText;
+    if (guard) {
+      getLogger().debug(`Guardrail '${guard.name}' triggered (pipeline)`);
+      text = guard.replacement ?? "I'm sorry, I can't respond to that.";
     }
-    const turn = this.metricsAcc.recordTurnComplete(responseText);
-    if (turn) {
-      this.deps.metricsStore.recordTurn({ call_id: this.callId, turn });
-      if (this.deps.onMetrics) await this.deps.onMetrics({ call_id: this.callId, turn });
+    this.metricsAcc.recordLlmComplete();
+    this.history.push({ role: "assistant", text, timestamp: Date.now() });
+    const chunker = new SentenceChunker();
+    const sentences = [...chunker.push(text), ...chunker.flush()];
+    const ttsFirstByteSent = { value: false };
+    this.beginSpeaking();
+    let interrupted = false;
+    try {
+      for (const sentence of sentences) {
+        if (!this.isSpeaking) {
+          interrupted = true;
+          break;
+        }
+        await this.synthesizeSentence(sentence, hookExecutor, hookCtx, ttsFirstByteSent);
+      }
+    } finally {
+      this.endSpeakingWithGrace();
     }
+    if (!interrupted) this.metricsAcc.recordTtsComplete(text);
+    return interrupted;
   }
   /** Handle streaming WebSocket remote response with TTS. */
   async handleWebSocketResponse(msgData) {
     const onMessage = this.deps.onMessage;
     const parts = [];
     this.metricsAcc.recordLlmComplete();
-    this.isSpeaking = true;
+    this.beginSpeaking();
     let wsTtsStarted = false;
     try {
       for await (const chunk of this.deps.remoteHandler.callWebSocket(onMessage, msgData)) {
         parts.push(chunk);
         if (this.tts) {
+          this.resetTtsCarry();
           for await (const audioChunk of this.tts.synthesizeStream(chunk)) {
             if (!this.isSpeaking) break;
             if (!wsTtsStarted) {
@@ -2879,15 +4869,12 @@ var StreamHandler = class {
     } catch (e) {
       getLogger().error(`WebSocket remote error (${this.deps.bridge.label}):`, e);
     } finally {
-      this.isSpeaking = false;
+      this.endSpeakingWithGrace();
+      this.resetTtsCarry();
     }
     const responseText = parts.join("");
     this.metricsAcc.recordTtsComplete(responseText);
-    const turn = this.metricsAcc.recordTurnComplete(responseText);
-    if (turn) {
-      this.deps.metricsStore.recordTurn({ call_id: this.callId, turn });
-      if (this.deps.onMetrics) await this.deps.onMetrics({ call_id: this.callId, turn });
-    }
+    await this.emitTurnMetrics(this.metricsAcc.recordTurnComplete(responseText));
     if (responseText) this.history.push({ role: "assistant", text: responseText, timestamp: Date.now() });
   }
   // ---------------------------------------------------------------------------
@@ -2922,86 +4909,92 @@ var StreamHandler = class {
     });
   }
   async handleAdapterEvent(type, eventData) {
-    if (type === "audio") {
-      if (!this.responseAudioStarted) {
-        this.responseAudioStarted = true;
-        if (this.metricsAcc.turnActive === false) {
-          this.metricsAcc.startTurn();
-        }
-        this.metricsAcc.recordTtsFirstByte();
-      }
-      let outAudio = eventData;
-      if (this.deps.bridge.telephonyProvider === "telnyx") {
-        outAudio = resample8kTo16k(mulawToPcm16(outAudio));
-      }
-      const encoded = outAudio.toString("base64");
-      this.deps.bridge.sendAudio(this.ws, encoded, this.streamSid);
-      this.chunkCount++;
-      this.deps.bridge.sendMark(this.ws, `audio_${this.chunkCount}`, this.streamSid);
-    } else if (type === "transcript_input") {
-      const inputText = eventData;
-      getLogger().debug(`User (${this.deps.bridge.label}): ${sanitizeLogValue(inputText)}`);
-      this.history.push({ role: "user", text: inputText, timestamp: Date.now() });
-      this.metricsAcc.startTurn();
-      this.currentAgentText = "";
-      this.responseAudioStarted = false;
-      if (this.deps.onTranscript) {
-        await this.deps.onTranscript({
-          role: "user",
-          text: inputText,
-          call_id: this.callId,
-          history: [...this.history.entries]
-        });
-      }
-    } else if (type === "transcript_output") {
-      const outputText = eventData;
-      if (outputText) {
-        const triggered = checkGuardrails(outputText, this.deps.agent.guardrails);
-        if (triggered) {
-          getLogger().debug(`Guardrail '${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.");
-          }
-        }
-        this.currentAgentText += outputText;
-      }
-    } else if (type === "response_done") {
-      const responseData = eventData;
-      if (responseData) {
-        const usage = responseData.usage;
-        if (usage) {
-          this.metricsAcc.recordRealtimeUsage(usage);
-        }
-      }
-      if (this.currentAgentText) {
-        this.history.push({ role: "assistant", text: this.currentAgentText, timestamp: Date.now() });
-        const turn = this.metricsAcc.recordTurnComplete(this.currentAgentText);
-        this.responseAudioStarted = false;
-        if (this.deps.onMetrics) {
-          await this.deps.onMetrics({
-            call_id: this.callId,
-            turn
-          });
-        }
-        this.deps.metricsStore.recordTurn({ call_id: this.callId, turn });
-        this.currentAgentText = "";
-      } else {
-        this.metricsAcc.recordTurnInterrupted();
-        this.responseAudioStarted = false;
+    const handler = this.adapterEventHandlers[type];
+    if (handler) await handler(eventData);
+  }
+  /** Event-type → handler dispatch table for the Realtime adapter. */
+  adapterEventHandlers = {
+    audio: async (eventData) => this.onAdapterAudio(eventData),
+    speech_stopped: async () => this.onAdapterSpeechStopped(),
+    transcript_input: async (eventData) => this.onAdapterTranscriptInput(eventData),
+    transcript_output: async (eventData) => this.onAdapterTranscriptOutput(eventData),
+    response_done: async (eventData) => this.onAdapterResponseDone(eventData),
+    speech_started: async () => this.onAdapterSpeechInterrupt(),
+    interruption: async () => this.onAdapterSpeechInterrupt(),
+    function_call: async (eventData) => {
+      if (this.adapter instanceof OpenAIRealtimeAdapter) {
+        await this.handleFunctionCall(eventData);
       }
-    } else if (type === "speech_started" || type === "interruption") {
-      this.deps.bridge.sendClear(this.ws, this.streamSid);
+    }
+  };
+  async onAdapterAudio(eventData) {
+    if (!this.responseAudioStarted) {
+      this.responseAudioStarted = true;
+      if (this.metricsAcc.turnActive === false) this.metricsAcc.startTurn();
+      this.metricsAcc.recordTtsFirstByte();
+    }
+    const outAudio = eventData;
+    this.deps.bridge.sendAudio(this.ws, outAudio.toString("base64"), this.streamSid);
+    this.chunkCount++;
+    this.deps.bridge.sendMark(this.ws, `audio_${this.chunkCount}`, this.streamSid);
+  }
+  onAdapterSpeechStopped() {
+    if (!this.metricsAcc.turnActive) this.metricsAcc.startTurn();
+    this.currentAgentText = "";
+    this.responseAudioStarted = false;
+  }
+  async onAdapterTranscriptInput(inputText) {
+    getLogger().debug(`User (${this.deps.bridge.label}): ${sanitizeLogValue(inputText)}`);
+    this.history.push({ role: "user", text: inputText, timestamp: Date.now() });
+    if (!this.metricsAcc.turnActive) {
+      this.metricsAcc.startTurn();
+      this.currentAgentText = "";
+      this.responseAudioStarted = false;
+    }
+    this.metricsAcc.recordSttComplete(inputText);
+    if (this.deps.onTranscript) {
+      await this.deps.onTranscript({
+        role: "user",
+        text: inputText,
+        call_id: this.callId,
+        history: [...this.history.entries]
+      });
+    }
+  }
+  async onAdapterTranscriptOutput(outputText) {
+    if (!outputText) return;
+    const triggered = checkGuardrails(outputText, this.deps.agent.guardrails);
+    if (triggered) {
+      getLogger().debug(`Guardrail '${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.");
       }
-      this.metricsAcc.recordTurnInterrupted();
+    }
+    this.currentAgentText += outputText;
+  }
+  async onAdapterResponseDone(responseData) {
+    if (responseData) {
+      const usage = responseData.usage;
+      if (usage) this.metricsAcc.recordRealtimeUsage(usage);
+    }
+    if (this.currentAgentText) {
+      this.history.push({ role: "assistant", text: this.currentAgentText, timestamp: Date.now() });
+      this.responseAudioStarted = false;
+      await this.emitTurnMetrics(this.metricsAcc.recordTurnComplete(this.currentAgentText));
       this.currentAgentText = "";
+    } else {
+      this.metricsAcc.recordTurnInterrupted();
       this.responseAudioStarted = false;
-    } else if (type === "function_call" && this.adapter instanceof OpenAIRealtimeAdapter) {
-      await this.handleFunctionCall(eventData);
     }
   }
+  onAdapterSpeechInterrupt() {
+    this.deps.bridge.sendClear(this.ws, this.streamSid);
+    if (this.adapter instanceof OpenAIRealtimeAdapter) this.adapter.cancelResponse();
+    this.metricsAcc.recordTurnInterrupted();
+    this.currentAgentText = "";
+    this.responseAudioStarted = false;
+  }
   async handleFunctionCall(fc) {
     const adapter = this.adapter;
     if (fc.name === "transfer_call") {
@@ -3095,7 +5088,7 @@ var StreamHandler = class {
       finalMetrics
     );
     try {
-      const { notifyDashboard } = await import("./persistence-CYIGNHSU.mjs");
+      const { notifyDashboard } = await import("./persistence-LQBYQPQQ.mjs");
       notifyDashboard(callEndData);
     } catch {
     }
@@ -3135,6 +5128,279 @@ async function queryDeepgramCost(metricsAcc, deepgramKey, deepgramRequestId) {
   }
 }
 
+// src/services/call-log.ts
+import * as crypto3 from "crypto";
+import * as fs2 from "fs";
+import { promises as fsp } from "fs";
+import * as os from "os";
+import * as path2 from "path";
+var SCHEMA_VERSION = "1.0";
+var DEFAULT_RETENTION_DAYS = 30;
+function xdgDataHome() {
+  return process.env.XDG_DATA_HOME || path2.join(os.homedir(), ".local", "share");
+}
+function platformDefaultRoot() {
+  if (process.platform === "darwin") {
+    return path2.join(os.homedir(), "Library", "Application Support", "patter");
+  }
+  if (process.platform === "win32") {
+    const localAppData = process.env.LOCALAPPDATA;
+    if (localAppData) return path2.join(localAppData, "patter");
+    return path2.join(os.homedir(), "AppData", "Local", "patter");
+  }
+  return path2.join(xdgDataHome(), "patter");
+}
+function resolveLogRoot(explicit) {
+  const value = explicit ?? process.env.PATTER_LOG_DIR;
+  if (!value) return null;
+  if (value.trim().toLowerCase() === "auto") return platformDefaultRoot();
+  if (value.startsWith("~")) return path2.join(os.homedir(), value.slice(1));
+  return value;
+}
+function retentionDays() {
+  const raw = process.env.PATTER_LOG_RETENTION_DAYS;
+  if (raw === void 0) return DEFAULT_RETENTION_DAYS;
+  const parsed = Number.parseInt(raw, 10);
+  if (Number.isNaN(parsed)) return DEFAULT_RETENTION_DAYS;
+  return Math.max(0, parsed);
+}
+function redactMode() {
+  const raw = (process.env.PATTER_LOG_REDACT_PHONE || "mask").trim().toLowerCase();
+  if (raw === "full" || raw === "mask" || raw === "hash_only") return raw;
+  return "mask";
+}
+function redactPhone(raw) {
+  if (!raw) return "";
+  const mode = redactMode();
+  if (mode === "full") return raw;
+  if (mode === "hash_only") {
+    return "sha256:" + crypto3.createHash("sha256").update(raw, "utf8").digest("hex").slice(0, 16);
+  }
+  return maskPhoneNumber(raw);
+}
+function utcIso(tsSeconds) {
+  const ms = tsSeconds !== void 0 ? tsSeconds * 1e3 : Date.now();
+  return new Date(ms).toISOString();
+}
+async function atomicWriteJson(filePath, payload) {
+  const dir = path2.dirname(filePath);
+  await fsp.mkdir(dir, { recursive: true });
+  const tmp = path2.join(dir, `.tmp.${process.pid}.${crypto3.randomBytes(4).toString("hex")}.json`);
+  try {
+    const handle = await fsp.open(tmp, "w");
+    try {
+      await handle.writeFile(JSON.stringify(payload, null, 2) + "\n", { encoding: "utf8" });
+      await handle.sync();
+    } finally {
+      await handle.close();
+    }
+    await fsp.rename(tmp, filePath);
+  } catch (err) {
+    try {
+      await fsp.unlink(tmp);
+    } catch {
+    }
+    throw err;
+  }
+}
+async function appendJsonl(filePath, record) {
+  await fsp.mkdir(path2.dirname(filePath), { recursive: true });
+  await fsp.appendFile(filePath, JSON.stringify(record) + "\n", { encoding: "utf8" });
+}
+var CallLogger = class {
+  root;
+  constructor(root) {
+    if (!root) {
+      this.root = null;
+      return;
+    }
+    const resolved = root.startsWith("~") ? path2.join(os.homedir(), root.slice(1)) : root;
+    try {
+      fs2.mkdirSync(resolved, { recursive: true });
+      this.root = resolved;
+      getLogger().info(`Call logs: ${resolved}`);
+    } catch (err) {
+      getLogger().warn(
+        `Could not create call log root ${resolved}: ${sanitizeLogValue(String(err))}`
+      );
+      this.root = null;
+    }
+  }
+  get enabled() {
+    return this.root !== null;
+  }
+  callDir(callId, startedAtSeconds) {
+    if (this.root === null) return null;
+    const ms = startedAtSeconds !== void 0 ? startedAtSeconds * 1e3 : Date.now();
+    const dt = new Date(ms);
+    const year = String(dt.getUTCFullYear()).padStart(4, "0");
+    const month = String(dt.getUTCMonth() + 1).padStart(2, "0");
+    const day = String(dt.getUTCDate()).padStart(2, "0");
+    const safeId = sanitizeLogValue(callId, 64).replace(/\//g, "_") || "unknown";
+    return path2.join(this.root, "calls", year, month, day, safeId);
+  }
+  async logCallStart(callId, input = {}) {
+    if (!this.enabled) return;
+    const startedAt = Date.now() / 1e3;
+    const dir = this.callDir(callId, startedAt);
+    if (dir === null) return;
+    const metadata = {
+      schema_version: SCHEMA_VERSION,
+      call_id: callId,
+      trace_id: input.traceId ?? null,
+      started_at: utcIso(startedAt),
+      ended_at: null,
+      duration_ms: null,
+      status: "in_progress",
+      caller: redactPhone(input.caller ?? ""),
+      callee: redactPhone(input.callee ?? ""),
+      telephony_provider: input.telephonyProvider ?? "",
+      provider_mode: input.providerMode ?? "",
+      agent: input.agent ?? {},
+      turns: 0,
+      cost: null,
+      latency: null,
+      error: null
+    };
+    try {
+      await atomicWriteJson(path2.join(dir, "metadata.json"), metadata);
+    } catch (err) {
+      getLogger().warn(`call_log write failed (${sanitizeLogValue(callId)}): ${sanitizeLogValue(String(err))}`);
+    }
+    if (crypto3.randomBytes(1)[0] < 5) {
+      this.sweepOldDays();
+    }
+  }
+  async logTurn(callId, turn) {
+    if (!this.enabled) return;
+    const dir = this.callDir(callId);
+    if (dir === null) return;
+    const record = {
+      schema_version: SCHEMA_VERSION,
+      ts: utcIso(typeof turn.timestamp === "number" ? turn.timestamp : void 0),
+      ...turn
+    };
+    try {
+      await appendJsonl(path2.join(dir, "transcript.jsonl"), record);
+    } catch (err) {
+      getLogger().warn(
+        `call_log turn write failed (${sanitizeLogValue(callId)}): ${sanitizeLogValue(String(err))}`
+      );
+    }
+  }
+  async logEvent(callId, eventType, payload = {}) {
+    if (!this.enabled) return;
+    const dir = this.callDir(callId);
+    if (dir === null) return;
+    const record = {
+      schema_version: SCHEMA_VERSION,
+      ts: utcIso(),
+      type: eventType,
+      data: payload
+    };
+    try {
+      await appendJsonl(path2.join(dir, "events.jsonl"), record);
+    } catch (err) {
+      getLogger().warn(
+        `call_log event write failed (${sanitizeLogValue(callId)}): ${sanitizeLogValue(String(err))}`
+      );
+    }
+  }
+  async logCallEnd(callId, input = {}) {
+    if (!this.enabled) return;
+    const dir = this.callDir(callId);
+    if (dir === null) return;
+    const metadataPath = path2.join(dir, "metadata.json");
+    let existing = {};
+    try {
+      existing = JSON.parse(await fsp.readFile(metadataPath, "utf8"));
+    } catch {
+      existing = {
+        schema_version: SCHEMA_VERSION,
+        call_id: callId,
+        started_at: null
+      };
+    }
+    const merged = {
+      ...existing,
+      ended_at: utcIso(),
+      duration_ms: input.durationSeconds !== void 0 ? Math.round(input.durationSeconds * 1e3 * 10) / 10 : null,
+      status: input.status ?? "completed",
+      turns: input.turns ?? null,
+      cost: input.cost ?? null,
+      latency: input.latency ?? null,
+      error: input.error ?? null
+    };
+    try {
+      await atomicWriteJson(metadataPath, merged);
+    } catch (err) {
+      getLogger().warn(
+        `call_log finalize failed (${sanitizeLogValue(callId)}): ${sanitizeLogValue(String(err))}`
+      );
+    }
+  }
+  // --- Retention ---------------------------------------------------------
+  sweepOldDays() {
+    if (this.root === null) return;
+    const days = retentionDays();
+    if (days === 0) return;
+    const cutoff = Date.now() / 1e3 - days * 86400;
+    const callsRoot = path2.join(this.root, "calls");
+    if (!fs2.existsSync(callsRoot)) return;
+    try {
+      for (const yearName of fs2.readdirSync(callsRoot)) {
+        if (!/^\d+$/.test(yearName)) continue;
+        const yearDir = path2.join(callsRoot, yearName);
+        if (!fs2.statSync(yearDir).isDirectory()) continue;
+        for (const monthName of fs2.readdirSync(yearDir)) {
+          if (!/^\d+$/.test(monthName)) continue;
+          const monthDir = path2.join(yearDir, monthName);
+          if (!fs2.statSync(monthDir).isDirectory()) continue;
+          for (const dayName of fs2.readdirSync(monthDir)) {
+            if (!/^\d+$/.test(dayName)) continue;
+            const dayDir = path2.join(monthDir, dayName);
+            const y = Number.parseInt(yearName, 10);
+            const m = Number.parseInt(monthName, 10);
+            const d = Number.parseInt(dayName, 10);
+            const ts = Date.UTC(y, m - 1, d) / 1e3;
+            if (ts < cutoff) {
+              rmTree(dayDir);
+            }
+          }
+          try {
+            if (fs2.readdirSync(monthDir).length === 0) fs2.rmdirSync(monthDir);
+          } catch {
+          }
+        }
+        try {
+          if (fs2.readdirSync(yearDir).length === 0) fs2.rmdirSync(yearDir);
+        } catch {
+        }
+      }
+    } catch (err) {
+      getLogger().debug(`call_log sweep failed: ${sanitizeLogValue(String(err))}`);
+    }
+  }
+};
+function rmTree(target) {
+  try {
+    for (const child of fs2.readdirSync(target)) {
+      const childPath = path2.join(target, child);
+      const stat = fs2.lstatSync(childPath);
+      if (stat.isDirectory()) {
+        rmTree(childPath);
+      } else {
+        try {
+          fs2.unlinkSync(childPath);
+        } catch {
+        }
+      }
+    }
+    fs2.rmdirSync(target);
+  } catch {
+  }
+}
+
 // src/server.ts
 var TRANSFER_CALL_TOOL = {
   name: "transfer_call",
@@ -3171,37 +5437,76 @@ function validateWebhookUrl(url) {
   if (!["http:", "https:"].includes(parsed.protocol)) {
     throw new Error(`Invalid webhook URL scheme: ${parsed.protocol}`);
   }
-  const hostname = parsed.hostname;
-  const blocked = [
-    /^127\./,
-    /^10\./,
-    /^172\.(1[6-9]|2\d|3[01])\./,
-    /^192\.168\./,
-    /^169\.254\./,
-    /^0\./,
-    /^::1$/,
-    /^localhost$/i,
-    /^metadata\.google\.internal$/i
-  ];
-  if (blocked.some((re) => re.test(hostname))) {
-    throw new Error(`Webhook URL blocked: ${hostname} is a private/internal address`);
+  const rawHost = parsed.hostname;
+  const host = rawHost.replace(/^\[/, "").replace(/\]$/, "").toLowerCase();
+  const BLOCKED_HOSTNAMES = /* @__PURE__ */ new Set([
+    "localhost",
+    "ip6-localhost",
+    "ip6-loopback",
+    "metadata",
+    "metadata.google.internal",
+    "metadata.azure.com"
+  ]);
+  if (BLOCKED_HOSTNAMES.has(host)) {
+    throw new Error(`Webhook URL blocked: ${rawHost} is a private/internal address`);
+  }
+  const IPV4_RE = /^(\d{1,3})\.(\d{1,3})\.(\d{1,3})\.(\d{1,3})$/;
+  const v4 = IPV4_RE.exec(host);
+  if (v4) {
+    const oct = v4.slice(1, 5).map((s) => parseInt(s, 10));
+    if (oct.some((n) => n < 0 || n > 255)) {
+      throw new Error(`Webhook URL blocked: ${rawHost} is not a valid IPv4 address`);
+    }
+    const [a, b] = oct;
+    if (a === 0 || // 0.0.0.0/8 (any 0.x)
+    a === 10 || // 10.0.0.0/8
+    a === 127 || // 127.0.0.0/8 loopback
+    a === 169 && b === 254 || // 169.254.0.0/16 link-local
+    a === 172 && b >= 16 && b <= 31 || // 172.16.0.0/12
+    a === 192 && b === 168) {
+      throw new Error(`Webhook URL blocked: ${rawHost} is a private/internal address`);
+    }
+    return;
+  }
+  if (host.includes(":")) {
+    if (host === "::1" || host === "::") {
+      throw new Error(`Webhook URL blocked: ${rawHost} is a private/internal address`);
+    }
+    if (/^fc[0-9a-f]{0,2}:/.test(host) || /^fd[0-9a-f]{0,2}:/.test(host)) {
+      throw new Error(`Webhook URL blocked: ${rawHost} is a private/internal address`);
+    }
+    if (/^fe[89ab][0-9a-f]?:/.test(host)) {
+      throw new Error(`Webhook URL blocked: ${rawHost} is a private/internal address`);
+    }
   }
 }
 function validateTelnyxSignature(rawBody, signature, timestamp, publicKey, toleranceSec = 300) {
   try {
     const ts = parseInt(timestamp, 10);
     if (!Number.isFinite(ts)) return false;
-    const ageMs = Date.now() - ts;
+    const tsMs = ts < 1e12 ? ts * 1e3 : ts;
+    const ageMs = Date.now() - tsMs;
     if (ageMs < 0 || ageMs > toleranceSec * 1e3) return false;
     const payload = `${timestamp}|${rawBody}`;
     const keyBuffer = Buffer.from(publicKey, "base64");
-    const sigBuffer = Buffer.from(signature, "base64");
-    const keyObject = crypto3.createPublicKey({
+    const keyObject = crypto4.createPublicKey({
       key: keyBuffer,
       format: "der",
       type: "spki"
     });
-    return crypto3.verify(null, Buffer.from(payload), keyObject, sigBuffer);
+    for (const rawSig of signature.split(",")) {
+      const trimmed = rawSig.trim();
+      if (!trimmed) continue;
+      try {
+        const sigBuffer = Buffer.from(trimmed, "base64");
+        if (crypto4.verify(null, Buffer.from(payload), keyObject, sigBuffer)) {
+          return true;
+        }
+      } catch {
+        continue;
+      }
+    }
+    return false;
   } catch {
     return false;
   }
@@ -3211,9 +5516,12 @@ function validateTwilioSid(sid, prefix = "CA") {
 }
 function validateTwilioSignature(url, params, signature, authToken) {
   const data = url + Object.keys(params).sort().reduce((acc, key) => acc + key + (params[key] ?? ""), "");
-  const expected = crypto3.createHmac("sha1", authToken).update(data).digest("base64");
+  const expected = crypto4.createHmac("sha1", authToken).update(data).digest("base64");
   try {
-    return crypto3.timingSafeEqual(Buffer.from(signature), Buffer.from(expected));
+    const sigBuf = Buffer.from(signature);
+    const expBuf = Buffer.from(expected);
+    if (sigBuf.length !== expBuf.length) return false;
+    return crypto4.timingSafeEqual(sigBuf, expBuf);
   } catch {
     return false;
   }
@@ -3246,9 +5554,7 @@ function buildAIAdapter(config, agent, resolvedPrompt) {
     return new ElevenLabsConvAIAdapter(
       engine.apiKey,
       engine.agentId,
-      agent.voice ?? "21m00Tcm4TlvDq8ikWAM",
-      "eleven_turbo_v2_5",
-      agent.language ?? "en",
+      agent.voice ?? "EXAVITQu4vr4xnSDxMaL",
       agent.firstMessage ?? ""
     );
   }
@@ -3353,7 +5659,7 @@ function isValidTelnyxTransferTarget(target) {
   if (/^\+[1-9]\d{6,14}$/.test(target)) return true;
   return /^sips?:[^\s@]+(@[^\s]+)?$/i.test(target);
 }
-var TELNYX_DTMF_ALLOWED = new Set("0123456789*#ABCDabcd");
+var TELNYX_DTMF_ALLOWED = new Set("0123456789*#ABCDabcdwW");
 var TELNYX_DTMF_DURATION_MS = 250;
 async function sleep(ms) {
   if (ms <= 0) return;
@@ -3379,7 +5685,7 @@ var TelnyxBridge = class {
       return;
     }
     const telnyxKey = this.config.telnyxKey ?? "";
-    await fetch(`https://api.telnyx.com/v2/calls/${callId}/actions/transfer`, {
+    await fetch(`https://api.telnyx.com/v2/calls/${encodeURIComponent(callId)}/actions/transfer`, {
       method: "POST",
       headers: { "Content-Type": "application/json", "Authorization": `Bearer ${telnyxKey}` },
       body: JSON.stringify({ to: toNumber })
@@ -3403,7 +5709,7 @@ var TelnyxBridge = class {
     }
     const duration = Math.max(100, Math.min(500, TELNYX_DTMF_DURATION_MS));
     for (let i = 0; i < filtered.length; i += 1) {
-      await fetch(`https://api.telnyx.com/v2/calls/${callId}/actions/send_dtmf`, {
+      await fetch(`https://api.telnyx.com/v2/calls/${encodeURIComponent(callId)}/actions/send_dtmf`, {
         method: "POST",
         headers: { "Content-Type": "application/json", "Authorization": `Bearer ${telnyxKey}` },
         body: JSON.stringify({ digits: filtered[i], duration_millis: duration })
@@ -3418,7 +5724,7 @@ var TelnyxBridge = class {
     const telnyxKey = this.config.telnyxKey ?? "";
     if (!telnyxKey || !callId) return;
     try {
-      const resp = await fetch(`https://api.telnyx.com/v2/calls/${callId}/actions/record_start`, {
+      const resp = await fetch(`https://api.telnyx.com/v2/calls/${encodeURIComponent(callId)}/actions/record_start`, {
         method: "POST",
         headers: { "Content-Type": "application/json", "Authorization": `Bearer ${telnyxKey}` },
         body: JSON.stringify({ format: "mp3", channels: "single" })
@@ -3436,7 +5742,7 @@ var TelnyxBridge = class {
     const telnyxKey = this.config.telnyxKey ?? "";
     if (!telnyxKey || !callId) return;
     try {
-      const resp = await fetch(`https://api.telnyx.com/v2/calls/${callId}/actions/record_stop`, {
+      const resp = await fetch(`https://api.telnyx.com/v2/calls/${encodeURIComponent(callId)}/actions/record_stop`, {
         method: "POST",
         headers: { "Content-Type": "application/json", "Authorization": `Bearer ${telnyxKey}` },
         body: JSON.stringify({})
@@ -3450,11 +5756,11 @@ var TelnyxBridge = class {
       getLogger().warn(`Telnyx record_stop error: ${String(e)}`);
     }
   }
-  async endCall(callId, ws) {
+  async endCall(callId, _ws) {
     const telnyxKey = this.config.telnyxKey ?? "";
     if (callId && telnyxKey) {
       try {
-        await fetch(`https://api.telnyx.com/v2/calls/${callId}/actions/hangup`, {
+        await fetch(`https://api.telnyx.com/v2/calls/${encodeURIComponent(callId)}/actions/hangup`, {
           method: "POST",
           headers: { "Content-Type": "application/json", "Authorization": `Bearer ${telnyxKey}` },
           body: JSON.stringify({})
@@ -3462,7 +5768,6 @@ var TelnyxBridge = class {
       } catch {
       }
     }
-    ws.close();
   }
   createStt(agent) {
     return createSTT(agent);
@@ -3471,7 +5776,7 @@ var TelnyxBridge = class {
     if (this.config.telnyxKey && callId) {
       try {
         const resp = await fetch(
-          `https://api.telnyx.com/v2/calls/${callId}`,
+          `https://api.telnyx.com/v2/calls/${encodeURIComponent(callId)}`,
           {
             headers: { "Authorization": `Bearer ${this.config.telnyxKey}` },
             signal: AbortSignal.timeout(5e3)
@@ -3506,6 +5811,17 @@ var EmbeddedServer = class {
     this.dashboardToken = dashboardToken;
     this.metricsStore = new MetricsStore();
     this.pricing = mergePricing(pricingOverrides);
+    const logRoot = resolveLogRoot();
+    if (logRoot) {
+      try {
+        const restored = this.metricsStore.hydrate(logRoot);
+        if (restored > 0) {
+          getLogger().info(`Dashboard hydrated ${restored} call(s) from ${logRoot}`);
+        }
+      } catch (err) {
+        getLogger().warn(`Dashboard hydration failed: ${String(err)}`);
+      }
+    }
   }
   server = null;
   wss = null;
@@ -3514,6 +5830,8 @@ var EmbeddedServer = class {
   metricsStore;
   pricing;
   remoteHandler = new RemoteMessageHandler();
+  /** Opt-in per-call filesystem logger (set via PATTER_LOG_DIR). */
+  callLogger = new CallLogger(resolveLogRoot());
   /** Active WebSocket connections tracked for graceful shutdown. */
   activeConnections = /* @__PURE__ */ new Set();
   activeCallIds = /* @__PURE__ */ new Map();
@@ -3522,6 +5840,18 @@ var EmbeddedServer = class {
     if (!webhookUrlPattern.test(this.config.webhookUrl)) {
       throw new Error(`Invalid webhookUrl: must be a hostname with no protocol prefix or path (got: '${this.config.webhookUrl}')`);
     }
+    if (this.config.requireSignature !== false) {
+      if (this.config.telephonyProvider === "twilio" && !this.config.twilioToken) {
+        getLogger().warn(
+          "Twilio webhook enforcement ACTIVE but twilioToken is empty \u2014 webhooks will 503. Set requireSignature=false for local dev."
+        );
+      }
+      if (this.config.telephonyProvider === "telnyx" && !this.config.telnyxPublicKey) {
+        getLogger().warn(
+          "Telnyx webhook enforcement ACTIVE but telnyxPublicKey is empty \u2014 webhooks will 503. Set requireSignature=false for local dev."
+        );
+      }
+    }
     const app = express();
     app.use((req, _res, next) => {
       if (req.path === "/webhooks/telnyx/voice") {
@@ -3549,14 +5879,8 @@ var EmbeddedServer = class {
       res.json({ status: "ok", mode: "local" });
     });
     if (this.dashboard) {
-      if (!this.dashboardToken) {
-        getLogger().warn(
-          "Dashboard is enabled without authentication. Set dashboardToken to protect call data. This is safe for local development but should not be exposed on a public network."
-        );
-      }
       mountDashboard(app, this.metricsStore, this.dashboardToken);
       mountApi(app, this.metricsStore, this.dashboardToken);
-      getLogger().info("Dashboard: http://127.0.0.1:" + port + "/");
     }
     app.post("/webhooks/twilio/status", (req, res) => {
       if (this.config.twilioToken) {
@@ -3567,6 +5891,10 @@ var EmbeddedServer = class {
           res.status(403).send("Invalid signature");
           return;
         }
+      } else if (this.config.requireSignature !== false) {
+        getLogger().error("Twilio webhook rejected: twilioToken not configured and requireSignature is not false");
+        res.status(503).send("Webhook signature required");
+        return;
       }
       const body = req.body;
       const callSid = sanitizeLogValue(body["CallSid"] ?? "");
@@ -3592,6 +5920,10 @@ var EmbeddedServer = class {
           res.status(403).send("Invalid signature");
           return;
         }
+      } else if (this.config.requireSignature !== false) {
+        getLogger().error("Twilio webhook rejected: twilioToken not configured and requireSignature is not false");
+        res.status(503).send("Webhook signature required");
+        return;
       }
       const body = req.body;
       const recordingSid = sanitizeLogValue(body["RecordingSid"] ?? "");
@@ -3609,6 +5941,10 @@ var EmbeddedServer = class {
           res.status(403).send("Invalid signature");
           return;
         }
+      } else if (this.config.requireSignature !== false) {
+        getLogger().error("Twilio webhook rejected: twilioToken not configured and requireSignature is not false");
+        res.status(503).send("Webhook signature required");
+        return;
       }
       const body = req.body;
       const answeredBy = body["AnsweredBy"] ?? "";
@@ -3651,6 +5987,10 @@ var EmbeddedServer = class {
           res.status(403).send("Invalid signature");
           return;
         }
+      } else if (this.config.requireSignature !== false) {
+        getLogger().error("Twilio webhook rejected: twilioToken not configured and requireSignature is not false");
+        res.status(503).send("Webhook signature required");
+        return;
       } else if (!this.twilioTokenWarningLogged) {
         this.twilioTokenWarningLogged = true;
         getLogger().warn("Twilio webhook signature validation disabled \u2014 set twilioToken for production");
@@ -3677,6 +6017,9 @@ var EmbeddedServer = class {
           getLogger().warn("Telnyx webhook rejected: invalid or missing Ed25519 signature");
           return res.status(403).send("Invalid signature");
         }
+      } else if (this.config.requireSignature !== false) {
+        getLogger().error("Telnyx webhook rejected: telnyxPublicKey not configured and requireSignature is not false");
+        return res.status(503).send("Webhook signature required");
       } else if (!this.telnyxSigWarningLogged) {
         this.telnyxSigWarningLogged = true;
         getLogger().warn("Telnyx webhook signature verification is disabled. Set telnyxPublicKey in LocalOptions for production use.");
@@ -3704,6 +6047,17 @@ var EmbeddedServer = class {
         }
         return res.status(200).send();
       }
+      if (eventType === "call.machine.detection.ended") {
+        const amdCallId = payload.call_control_id ?? "";
+        const amdResult = String(payload.result ?? "");
+        getLogger().info(
+          `Telnyx AMD result for ${sanitizeLogValue(amdCallId)}: ${sanitizeLogValue(amdResult)}`
+        );
+        if (amdCallId && (amdResult === "machine" || amdResult === "machine_detected")) {
+          await this.handleTelnyxAmdVoicemail(amdCallId);
+        }
+        return res.status(200).send();
+      }
       const callControlId = payload.call_control_id ?? "";
       if (!callControlId) {
         getLogger().warn("Telnyx webhook rejected: missing call_control_id");
@@ -3721,27 +6075,18 @@ var EmbeddedServer = class {
       };
       try {
         if (eventType === "call.initiated") {
-          getLogger().info(`Telnyx call.initiated ${callControlId} \u2014 answering`);
-          const resp = await fetch(`${apiBase}/calls/${encodeURIComponent(callControlId)}/actions/answer`, {
-            method: "POST",
-            headers: authHeaders,
-            body: JSON.stringify({}),
-            signal: AbortSignal.timeout(1e4)
-          });
-          if (!resp.ok) {
-            getLogger().warn(`Telnyx answer failed: ${resp.status} ${(await resp.text()).slice(0, 200)}`);
-          }
-        } else if (eventType === "call.answered") {
           const caller = payload.from ?? "";
           const callee = payload.to ?? "";
           const streamUrl = `wss://${this.config.webhookUrl}/ws/stream/${encodeURIComponent(callControlId)}?caller=${encodeURIComponent(caller)}&callee=${encodeURIComponent(callee)}`;
-          getLogger().info(`Telnyx call.answered ${callControlId} \u2014 starting stream`);
-          const resp = await fetch(`${apiBase}/calls/${encodeURIComponent(callControlId)}/actions/streaming_start`, {
+          getLogger().info(`Telnyx call.initiated ${callControlId} \u2014 answering with inline stream`);
+          const resp = await fetch(`${apiBase}/calls/${encodeURIComponent(callControlId)}/actions/answer`, {
             method: "POST",
             headers: authHeaders,
             body: JSON.stringify({
               stream_url: streamUrl,
-              stream_track: "both_tracks",
+              // ``inbound_track`` halves WS upstream bandwidth — outbound
+              // echo was always filtered downstream anyway.
+              stream_track: "inbound_track",
               stream_bidirectional_mode: "rtp",
               stream_bidirectional_codec: "PCMU",
               stream_bidirectional_sampling_rate: 8e3,
@@ -3750,8 +6095,10 @@ var EmbeddedServer = class {
             signal: AbortSignal.timeout(1e4)
           });
           if (!resp.ok) {
-            getLogger().warn(`Telnyx streaming_start failed: ${resp.status} ${(await resp.text()).slice(0, 200)}`);
+            getLogger().warn(`Telnyx answer failed: ${resp.status} ${(await resp.text()).slice(0, 200)}`);
           }
+        } else if (eventType === "call.answered") {
+          getLogger().debug(`Telnyx call.answered ${callControlId} \u2014 stream already active (inline)`);
         } else {
           getLogger().debug(`Telnyx event ignored: ${eventType}`);
         }
@@ -3773,7 +6120,6 @@ var EmbeddedServer = class {
         socket.destroy();
         return;
       }
-      getLogger().info(`Upgrade request: ${req.url}`);
       this.wss.handleUpgrade(req, socket, head, (ws) => {
         wsConnectionsByIp.set(remoteIp, (wsConnectionsByIp.get(remoteIp) ?? 0) + 1);
         ws.once("close", () => {
@@ -3789,7 +6135,6 @@ var EmbeddedServer = class {
     });
     this.wss.on("connection", (ws, req) => {
       const url = new URL(req.url ?? "", `http://localhost`);
-      getLogger().info(`WebSocket connected: ${req.url}`);
       this.activeConnections.add(ws);
       ws.once("close", () => {
         this.activeConnections.delete(ws);
@@ -3803,28 +6148,86 @@ var EmbeddedServer = class {
     });
     await new Promise((resolve) => {
       this.server.listen(port, "127.0.0.1", () => {
-        getLogger().info(`
-\u2588\u2588\u2588\u2588\u2588\u2588\u2557  \u2588\u2588\u2588\u2588\u2588\u2557 \u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2588\u2588\u2557
-\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557\u255A\u2550\u2550\u2588\u2588\u2554\u2550\u2550\u255D\u255A\u2550\u2550\u2588\u2588\u2554\u2550\u2550\u255D\u2588\u2588\u2554\u2550\u2550\u2550\u2550\u255D\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557
-\u2588\u2588\u2588\u2588\u2588\u2588\u2554\u255D\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2551   \u2588\u2588\u2551      \u2588\u2588\u2551   \u2588\u2588\u2588\u2588\u2588\u2557  \u2588\u2588\u2588\u2588\u2588\u2588\u2554\u255D
-\u2588\u2588\u2554\u2550\u2550\u2550\u255D \u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2551   \u2588\u2588\u2551      \u2588\u2588\u2551   \u2588\u2588\u2554\u2550\u2550\u255D  \u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557
-\u2588\u2588\u2551     \u2588\u2588\u2551  \u2588\u2588\u2551   \u2588\u2588\u2551      \u2588\u2588\u2551   \u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2551  \u2588\u2588\u2551
-\u255A\u2550\u255D     \u255A\u2550\u255D  \u255A\u2550\u255D   \u255A\u2550\u255D      \u255A\u2550\u255D   \u255A\u2550\u2550\u2550\u2550\u2550\u2550\u255D\u255A\u2550\u255D  \u255A\u2550\u255D
-
-Connect AI agents to phone numbers in 4 lines of code
-`);
         getLogger().info(`Server on port ${port}`);
         getLogger().info(`Webhook: https://${this.config.webhookUrl}`);
-        getLogger().info(`Phone: ${this.config.phoneNumber}`);
+        getLogger().info(`Phone:   ${this.config.phoneNumber}`);
+        const model = this.agent.model ?? "";
+        if (model && model !== "gpt-4o-mini-realtime-preview" && model.includes("realtime")) {
+          getLogger().warn(
+            `Agent uses "${sanitizeLogValue(model)}" but DEFAULT_PRICING.openai_realtime is calibrated for "gpt-4o-mini-realtime-preview". Pass Patter({ pricing: { openai_realtime: {...} } }) to set rates for this model, otherwise the dashboard cost display will under-report.`
+          );
+        }
+        if (this.dashboard) {
+          console.log("\n\u2500\u2500\u2500\u2500 Dashboard \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500");
+          getLogger().info(`URL: http://127.0.0.1:${port}/`);
+          if (!this.dashboardToken) {
+            getLogger().warn(
+              "Dashboard is enabled without authentication. Set dashboardToken to protect call data. This is safe for local development but should not be exposed on a public network."
+            );
+          }
+          console.log("\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n");
+        }
         resolve();
       });
     });
   }
+  /**
+   * Handle a Telnyx ``call.machine.detection.ended`` event when AMD returns
+   * ``machine``: speak the configured voicemail message via ``actions/speak``
+   * then hang up via ``actions/hangup``. Mirrors the Python
+   * ``handle_amd_result`` helper.
+   */
+  async handleTelnyxAmdVoicemail(callControlId) {
+    const telnyxKey = this.config.telnyxKey ?? "";
+    if (!callControlId || !telnyxKey || !this.voicemailMessage) {
+      return;
+    }
+    const encoded = encodeURIComponent(callControlId);
+    const headers = {
+      "Content-Type": "application/json",
+      Authorization: `Bearer ${telnyxKey}`
+    };
+    const estimatedMs = Math.min(
+      3e4,
+      Math.ceil(this.voicemailMessage.length / 14 * 1e3) + 1500
+    );
+    try {
+      const speakResp = await fetch(
+        `https://api.telnyx.com/v2/calls/${encoded}/actions/speak`,
+        {
+          method: "POST",
+          headers,
+          body: JSON.stringify({
+            payload: this.voicemailMessage,
+            voice: "female",
+            language: "en-US"
+          }),
+          signal: AbortSignal.timeout(1e4)
+        }
+      );
+      if (!speakResp.ok) {
+        getLogger().warn(
+          `Telnyx voicemail speak failed: ${speakResp.status} ${(await speakResp.text()).slice(0, 200)}`
+        );
+      }
+      await new Promise((resolve) => setTimeout(resolve, estimatedMs));
+      await fetch(`https://api.telnyx.com/v2/calls/${encoded}/actions/hangup`, {
+        method: "POST",
+        headers,
+        body: JSON.stringify({}),
+        signal: AbortSignal.timeout(1e4)
+      });
+      getLogger().info(`Voicemail dropped for Telnyx call ${sanitizeLogValue(callControlId)}`);
+    } catch (e) {
+      getLogger().warn(`Could not drop voicemail (Telnyx): ${String(e)}`);
+    }
+  }
   // ---------------------------------------------------------------------------
   // Stream handler helpers
   // ---------------------------------------------------------------------------
   /** Build the shared StreamHandlerDeps for the current server configuration. */
   buildStreamHandlerDeps(bridge) {
+    const [wrappedStart, wrappedMetrics, wrappedEnd] = this.wrapLoggingCallbacks(bridge);
     return {
       config: this.config,
       agent: this.agent,
@@ -3832,17 +6235,84 @@ Connect AI agents to phone numbers in 4 lines of code
       metricsStore: this.metricsStore,
       pricing: this.pricing,
       remoteHandler: this.remoteHandler,
-      onCallStart: this.onCallStart,
-      onCallEnd: this.onCallEnd,
+      onCallStart: wrappedStart,
+      onCallEnd: wrappedEnd,
       onTranscript: this.onTranscript,
       onMessage: this.onMessage,
-      onMetrics: this.onMetrics,
+      onMetrics: wrappedMetrics,
       recording: this.recording,
       buildAIAdapter: (resolvedPrompt) => buildAIAdapter(this.config, this.agent, resolvedPrompt),
       sanitizeVariables,
       resolveVariables
     };
   }
+  /**
+   * Wrap user-supplied call lifecycle callbacks with CallLogger side-effects.
+   * When PATTER_LOG_DIR is unset, the logger is disabled and the returned
+   * wrappers degrade to just calling the user callbacks (still wrapped so
+   * the logger stays consistent with future configuration changes).
+   */
+  wrapLoggingCallbacks(bridge) {
+    const logger = this.callLogger;
+    const agent = this.agent;
+    const userStart = this.onCallStart;
+    const userMetrics = this.onMetrics;
+    const userEnd = this.onCallEnd;
+    const agentSnapshot = () => {
+      const snap = {
+        provider: agent.provider,
+        model: agent.model,
+        voice: agent.voice,
+        language: agent.language
+      };
+      if (agent.stt && agent.tts && !("engine" in agent && agent.engine)) {
+        snap.mode = "pipeline";
+      }
+      return Object.fromEntries(Object.entries(snap).filter(([, v]) => v !== void 0));
+    };
+    const wrappedStart = async (data) => {
+      if (logger.enabled) {
+        const callId = typeof data.call_id === "string" ? data.call_id : "";
+        void logger.logCallStart(callId, {
+          caller: typeof data.caller === "string" ? data.caller : "",
+          callee: typeof data.callee === "string" ? data.callee : "",
+          telephonyProvider: bridge.telephonyProvider,
+          providerMode: agent.provider ?? "",
+          agent: agentSnapshot()
+        }).catch((err) => getLogger().error(`call_log start error: ${String(err)}`));
+      }
+      if (userStart) await userStart(data);
+    };
+    const wrappedMetrics = async (data) => {
+      if (logger.enabled) {
+        const callId = typeof data.call_id === "string" ? data.call_id : "";
+        const turn = data.turn;
+        if (turn && typeof turn === "object") {
+          void logger.logTurn(callId, turn).catch((err) => getLogger().error(`call_log turn error: ${String(err)}`));
+        }
+      }
+      if (userMetrics) await userMetrics(data);
+    };
+    const wrappedEnd = async (data) => {
+      if (logger.enabled) {
+        const callId = typeof data.call_id === "string" ? data.call_id : "";
+        const metricsObj = data.metrics ?? null;
+        const latency = metricsObj ? {
+          p50_ms: metricsObj.latency_p50?.total_ms ?? null,
+          p95_ms: metricsObj.latency_p95?.total_ms ?? null,
+          p99_ms: metricsObj.latency_p99?.total_ms ?? null
+        } : null;
+        void logger.logCallEnd(callId, {
+          durationSeconds: metricsObj?.duration_seconds,
+          turns: metricsObj?.turns?.length,
+          cost: metricsObj?.cost ?? null,
+          latency
+        }).catch((err) => getLogger().error(`call_log end error: ${String(err)}`));
+      }
+      if (userEnd) await userEnd(data);
+    };
+    return [wrappedStart, wrappedMetrics, wrappedEnd];
+  }
   // ---------------------------------------------------------------------------
   // Twilio WebSocket message parser (thin layer)
   // ---------------------------------------------------------------------------
@@ -3861,7 +6331,6 @@ Connect AI agents to phone numbers in 4 lines of code
           return;
         }
         const event = data.event;
-        getLogger().info(`WS event: ${event}`);
         if (event === "start") {
           handler.setStreamSid(data.streamSid ?? "");
           const callSid = data.start?.callSid ?? "";
@@ -3872,6 +6341,8 @@ Connect AI agents to phone numbers in 4 lines of code
           const payload = data.media?.payload ?? "";
           handler.handleAudio(Buffer.from(payload, "base64"));
         } else if (event === "mark") {
+          const markName = String(data.mark?.name ?? "");
+          if (markName) await handler.onMark(markName);
         } else if (event === "dtmf") {
           const digit = data.dtmf?.digit ?? "";
           await handler.handleDtmf(digit);
@@ -3907,7 +6378,6 @@ Connect AI agents to phone numbers in 4 lines of code
         }
         const event = data.event ?? "";
         if (event === "connected") return;
-        getLogger().info(`Telnyx event: ${event}`);
         if (event === "start" && !streamStarted) {
           streamStarted = true;
           const callControlId = data.start?.call_control_id ?? "";
@@ -4008,19 +6478,145 @@ Connect AI agents to phone numbers in 4 lines of code
 };
 
 // src/llm-loop.ts
+var DEFAULT_TOOL_MAX_RETRIES = 2;
+var DEFAULT_TOOL_RETRY_DELAY_MS = 500;
+var DEFAULT_TOOL_TIMEOUT_MS = 1e4;
+var TOOL_MAX_RESPONSE_BYTES = 1 * 1024 * 1024;
+var DefaultToolExecutor = class {
+  maxRetries;
+  retryDelayMs;
+  requestTimeoutMs;
+  constructor(opts = {}) {
+    this.maxRetries = opts.maxRetries ?? DEFAULT_TOOL_MAX_RETRIES;
+    this.retryDelayMs = opts.retryDelayMs ?? DEFAULT_TOOL_RETRY_DELAY_MS;
+    this.requestTimeoutMs = opts.requestTimeoutMs ?? DEFAULT_TOOL_TIMEOUT_MS;
+  }
+  async execute(toolDef, args, callContext) {
+    if (toolDef.handler) {
+      try {
+        return await toolDef.handler(args, callContext);
+      } catch (e) {
+        return JSON.stringify({
+          error: `Tool handler error: ${String(e)}`,
+          fallback: true
+        });
+      }
+    }
+    if (toolDef.webhookUrl) {
+      try {
+        validateWebhookUrl(toolDef.webhookUrl);
+      } catch (e) {
+        return JSON.stringify({ error: `Tool webhook URL rejected: ${String(e)}` });
+      }
+      const callId = typeof callContext.call_id === "string" ? callContext.call_id : "";
+      return await withSpan(
+        SPAN_TOOL,
+        {
+          "patter.tool.name": toolDef.name,
+          "patter.tool.transport": "webhook",
+          "patter.call.id": callId
+        },
+        async (span) => {
+          const totalAttempts = this.maxRetries + 1;
+          for (let attempt = 0; attempt < totalAttempts; attempt++) {
+            span.setAttribute("patter.tool.attempt", attempt + 1);
+            try {
+              const resp = await fetch(toolDef.webhookUrl, {
+                method: "POST",
+                headers: { "Content-Type": "application/json" },
+                body: JSON.stringify({
+                  tool: toolDef.name,
+                  arguments: args,
+                  ...callContext,
+                  attempt: attempt + 1
+                }),
+                signal: AbortSignal.timeout(this.requestTimeoutMs)
+              });
+              if (!resp.ok) throw new Error(`HTTP ${resp.status}`);
+              const result = JSON.stringify(await resp.json());
+              if (result.length > TOOL_MAX_RESPONSE_BYTES) {
+                return JSON.stringify({
+                  error: `Webhook response too large: ${result.length} bytes (max ${TOOL_MAX_RESPONSE_BYTES})`,
+                  fallback: true
+                });
+              }
+              return result;
+            } catch (e) {
+              if (attempt < totalAttempts - 1) {
+                getLogger().warn(
+                  `Tool webhook '${toolDef.name}' failed (attempt ${attempt + 1}), retrying: ${String(e)}`
+                );
+                await new Promise((r) => setTimeout(r, this.retryDelayMs));
+              } else {
+                span.recordException(e);
+                return JSON.stringify({
+                  error: `Tool failed after ${totalAttempts} attempts: ${String(e)}`,
+                  fallback: true
+                });
+              }
+            }
+          }
+          return JSON.stringify({
+            error: `Tool '${toolDef.name}' exited retry loop unexpectedly`,
+            fallback: true
+          });
+        }
+      );
+    }
+    return JSON.stringify({
+      error: `No handler or webhookUrl for tool '${toolDef.name}'`,
+      fallback: true
+    });
+  }
+};
 var OpenAILLMProvider = class {
   apiKey;
   model;
-  constructor(apiKey, model) {
+  temperature;
+  maxTokens;
+  responseFormat;
+  parallelToolCalls;
+  toolChoice;
+  seed;
+  topP;
+  frequencyPenalty;
+  presencePenalty;
+  stop;
+  constructor(apiKey, model, sampling = {}) {
     this.apiKey = apiKey;
     this.model = model;
+    this.temperature = sampling.temperature;
+    this.maxTokens = sampling.maxTokens;
+    this.responseFormat = sampling.responseFormat;
+    this.parallelToolCalls = sampling.parallelToolCalls;
+    this.toolChoice = sampling.toolChoice;
+    this.seed = sampling.seed;
+    this.topP = sampling.topP;
+    this.frequencyPenalty = sampling.frequencyPenalty;
+    this.presencePenalty = sampling.presencePenalty;
+    this.stop = sampling.stop;
   }
   async *stream(messages, tools) {
     const body = {
       model: this.model,
       messages,
-      stream: true
+      stream: true,
+      // Ask OpenAI to include a final usage chunk so we can attribute token
+      // cost. Without this the dashboard shows LLM cost = 0 for OpenAI.
+      stream_options: { include_usage: true }
     };
+    if (this.temperature !== void 0) body.temperature = this.temperature;
+    if (this.maxTokens !== void 0) {
+      body.max_completion_tokens = this.maxTokens;
+    }
+    if (this.responseFormat !== void 0) body.response_format = this.responseFormat;
+    if (this.parallelToolCalls !== void 0) body.parallel_tool_calls = this.parallelToolCalls;
+    if (this.toolChoice !== void 0) body.tool_choice = this.toolChoice;
+    if (this.seed !== void 0) body.seed = this.seed;
+    if (this.topP !== void 0) body.top_p = this.topP;
+    if (this.frequencyPenalty !== void 0) body.frequency_penalty = this.frequencyPenalty;
+    if (this.presencePenalty !== void 0) body.presence_penalty = this.presencePenalty;
+    if (this.stop !== void 0) body.stop = this.stop;
     if (tools) {
       body.tools = tools;
     }
@@ -4059,6 +6655,16 @@ var OpenAILLMProvider = class {
         } catch {
           continue;
         }
+        if (chunk.usage) {
+          const cached = chunk.usage.prompt_tokens_details?.cached_tokens ?? 0;
+          const uncachedInput = Math.max(0, (chunk.usage.prompt_tokens ?? 0) - cached);
+          yield {
+            type: "usage",
+            inputTokens: uncachedInput,
+            outputTokens: chunk.usage.completion_tokens,
+            cacheReadInputTokens: cached
+          };
+        }
         const delta = chunk.choices?.[0]?.delta;
         if (!delta) continue;
         if (delta.content) {
@@ -4085,10 +6691,28 @@ var LLMLoop = class {
   tools;
   openaiTools;
   toolMap;
+  toolExecutor;
+  eventBus;
+  // Fix 10: track provider/model so usage chunks can be attributed for billing.
+  _providerName;
+  _modelName;
   constructor(apiKey, model, systemPrompt, tools, llmProvider) {
     this.provider = llmProvider ?? new OpenAILLMProvider(apiKey, model);
     this.systemPrompt = systemPrompt;
+    if (llmProvider) {
+      const key = llmProvider.constructor?.providerKey;
+      if (key) {
+        this._providerName = key;
+      } else {
+        const stripped = (llmProvider.constructor?.name ?? "custom").replace(/LLMProvider$/i, "").replace(/LLM$/i, "").replace(/Provider$/i, "").toLowerCase();
+        this._providerName = stripped || "custom";
+      }
+    } else {
+      this._providerName = "openai";
+    }
+    this._modelName = model;
     this.tools = tools ?? null;
+    this.toolExecutor = new DefaultToolExecutor();
     this.toolMap = /* @__PURE__ */ new Map();
     this.openaiTools = null;
     if (this.tools && this.tools.length > 0) {
@@ -4106,13 +6730,40 @@ var LLMLoop = class {
       }
     }
   }
+  /**
+   * Swap in a custom tool executor (e.g. different retry policy, metrics
+   * wrapping, tenant-aware fan-out). The default is ``DefaultToolExecutor``.
+   */
+  setToolExecutor(executor) {
+    this.toolExecutor = executor;
+  }
+  /**
+   * Wire an :class:`EventBus` so the loop emits ``llm_chunk`` per text
+   * token and ``tool_call_started`` the first time each tool-call index
+   * appears. Set to ``undefined`` to disable.
+   */
+  setEventBus(bus) {
+    this.eventBus = bus;
+  }
   /**
    * Stream LLM response tokens, handling tool calls automatically.
    * Yields text tokens as they arrive from the LLM.
+   *
+   * @param metrics Optional usage recorder — when provided, usage chunks
+   *   from the provider are forwarded to {@link LlmUsageRecorder.recordLlmUsage}
+   *   so token costs are included in the call cost breakdown (fix 10).
    */
-  async *run(userText, history, callContext) {
-    const messages = this.buildMessages(history, userText);
+  async *run(userText, history, callContext, metrics, hookExecutor, hookCtx) {
+    let messages = this.buildMessages(history, userText);
     const maxIterations = 10;
+    if (hookExecutor && hookCtx) {
+      messages = await hookExecutor.runBeforeLlm(
+        messages,
+        hookCtx
+      );
+    }
+    const hasAfterLlm = Boolean(hookExecutor?.hasAfterLlm() && hookCtx);
+    const allEmittedText = [];
     for (let iter = 0; iter < maxIterations; iter++) {
       const toolCallsAccumulated = /* @__PURE__ */ new Map();
       const textParts = [];
@@ -4120,12 +6771,31 @@ var LLMLoop = class {
       for await (const chunk of this.provider.stream(messages, this.openaiTools)) {
         if (chunk.type === "text" && chunk.content) {
           textParts.push(chunk.content);
-          yield chunk.content;
+          this.eventBus?.emit("llm_chunk", { text: chunk.content, iteration: iter });
+          if (hasAfterLlm) {
+            allEmittedText.push(chunk.content);
+          } else {
+            yield chunk.content;
+          }
+        } else if (chunk.type === "usage") {
+          metrics?.recordLlmUsage(
+            this._providerName,
+            this._modelName,
+            chunk.inputTokens ?? 0,
+            chunk.outputTokens ?? 0,
+            chunk.cacheReadInputTokens ?? 0,
+            chunk.cacheCreationInputTokens ?? 0
+          );
         } else if (chunk.type === "tool_call") {
           hasToolCalls = true;
           const idx = chunk.index ?? 0;
           if (!toolCallsAccumulated.has(idx)) {
             toolCallsAccumulated.set(idx, { id: "", name: "", arguments: "" });
+            this.eventBus?.emit("tool_call_started", {
+              index: idx,
+              name: chunk.name ?? "",
+              args: chunk.arguments ?? ""
+            });
           }
           const acc = toolCallsAccumulated.get(idx);
           if (chunk.id) acc.id = chunk.id;
@@ -4133,7 +6803,14 @@ var LLMLoop = class {
           if (chunk.arguments) acc.arguments += chunk.arguments;
         }
       }
-      if (!hasToolCalls) return;
+      if (!hasToolCalls) {
+        if (hasAfterLlm && hookExecutor && hookCtx) {
+          const finalText = allEmittedText.join("");
+          const rewritten = await hookExecutor.runAfterLlm(finalText, hookCtx);
+          if (rewritten) yield rewritten;
+        }
+        return;
+      }
       const assistantMsg = {
         role: "assistant",
         content: textParts.join("") || null,
@@ -4172,49 +6849,7 @@ var LLMLoop = class {
     if (!toolDef) {
       return JSON.stringify({ error: `Unknown tool: ${toolName}` });
     }
-    if (toolDef.handler) {
-      try {
-        return await toolDef.handler(args, callContext);
-      } catch (e) {
-        return JSON.stringify({ error: `Tool handler error: ${String(e)}` });
-      }
-    }
-    if (toolDef.webhookUrl) {
-      try {
-        validateWebhookUrl(toolDef.webhookUrl);
-      } catch (e) {
-        return JSON.stringify({ error: `Tool webhook URL rejected: ${String(e)}` });
-      }
-      for (let attempt = 0; attempt < 3; attempt++) {
-        try {
-          const resp = await fetch(toolDef.webhookUrl, {
-            method: "POST",
-            headers: { "Content-Type": "application/json" },
-            body: JSON.stringify({
-              tool: toolName,
-              arguments: args,
-              ...callContext,
-              attempt: attempt + 1
-            }),
-            signal: AbortSignal.timeout(1e4)
-          });
-          if (!resp.ok) throw new Error(`HTTP ${resp.status}`);
-          const result = JSON.stringify(await resp.json());
-          const MAX_RESPONSE_BYTES2 = 1 * 1024 * 1024;
-          if (result.length > MAX_RESPONSE_BYTES2) {
-            return JSON.stringify({ error: `Webhook response too large: ${result.length} bytes (max ${MAX_RESPONSE_BYTES2})`, fallback: true });
-          }
-          return result;
-        } catch (e) {
-          if (attempt < 2) {
-            await new Promise((r) => setTimeout(r, 500));
-          } else {
-            return JSON.stringify({ error: `Tool failed after 3 attempts: ${String(e)}` });
-          }
-        }
-      }
-    }
-    return JSON.stringify({ error: `No handler or webhookUrl for tool '${toolName}'` });
+    return this.toolExecutor.execute(toolDef, args, callContext);
   }
   buildMessages(history, userText) {
     const messages = [
@@ -4401,6 +7036,11 @@ var TestSession = class {
 };
 
 export {
+  PatterError,
+  PatterConnectionError,
+  AuthenticationError,
+  ProvisionError,
+  RateLimitError,
   OpenAIRealtimeAdapter,
   ElevenLabsConvAIAdapter,
   DEFAULT_PRICING,
@@ -4422,14 +7062,31 @@ export {
   CallMetricsAccumulator,
   mulawToPcm16,
   pcm16ToMulaw,
+  PcmCarry,
+  StatefulResampler,
+  createResampler16kTo8k,
+  createResampler8kTo16k,
+  createResampler24kTo16k,
   resample8kTo16k,
   resample16kTo8k,
   resample24kTo16k,
+  SPAN_CALL,
+  SPAN_STT,
+  SPAN_LLM,
+  SPAN_TTS,
+  SPAN_TOOL,
+  SPAN_ENDPOINT,
+  SPAN_BARGEIN,
+  initTracing,
+  isTracingEnabled,
+  startSpan,
+  DefaultToolExecutor,
   OpenAILLMProvider,
   LLMLoop,
   DEFAULT_MIN_SENTENCE_LEN,
   SentenceChunker,
   PipelineHookExecutor,
+  EventBus,
   EmbeddedServer,
   TestSession
 };