realtime-voice-agents 2.0.0

package/dist/mulaw--cwU2c9L.mjs

@@ -0,0 +1,64 @@
+//#region src/audio/mulaw.ts
+/**
+* G.711 μ-law codec (the format Twilio Media Streams speak: 8 kHz, 1 byte/sample).
+*
+* Decode uses a 256-entry lookup table; encode uses the canonical Sun g711.c
+* algorithm (bias 0x84, clip 32635). μ-law silence is 0xFF.
+*/
+/** One byte of μ-law silence (encodes linear 0). */
+const MULAW_SILENCE_BYTE = 255;
+/** Samples per second on the Twilio media stream. μ-law is 1 byte per sample. */
+const MULAW_SAMPLE_RATE = 8e3;
+/** Bytes per 20 ms Twilio media frame at 8 kHz μ-law. */
+const MULAW_FRAME_BYTES_20MS = 160;
+function decodeSampleUncached(byte) {
+	const u = ~byte & 255;
+	let t = ((u & 15) << 3) + 132;
+	t <<= (u & 112) >> 4;
+	return u & 128 ? 132 - t : t - 132;
+}
+const DECODE_TABLE = /* @__PURE__ */ new Int16Array(256);
+for (let i = 0; i < 256; i++) DECODE_TABLE[i] = decodeSampleUncached(i);
+const BIAS = 132;
+const CLIP = 32635;
+/** Decode a single μ-law byte to a linear PCM16 sample. */
+function mulawDecodeSample(byte) {
+	return DECODE_TABLE[byte & 255];
+}
+/** Encode a single linear PCM16 sample to a μ-law byte. */
+function mulawEncodeSample(sample) {
+	let pcm = sample;
+	const sign = pcm < 0 ? 128 : 0;
+	if (sign) pcm = -pcm;
+	if (pcm > CLIP) pcm = CLIP;
+	pcm += BIAS;
+	let exponent = 7;
+	for (let mask = 16384; (pcm & mask) === 0 && exponent > 0; mask >>= 1) exponent--;
+	const mantissa = pcm >> exponent + 3 & 15;
+	return ~(sign | exponent << 4 | mantissa) & 255;
+}
+/** Decode a μ-law buffer to linear PCM16 samples. */
+function mulawToPcm16(mulaw) {
+	const out = new Int16Array(mulaw.length);
+	for (let i = 0; i < mulaw.length; i++) out[i] = DECODE_TABLE[mulaw[i]];
+	return out;
+}
+/** Encode linear PCM16 samples to a μ-law buffer. */
+function pcm16ToMulaw(pcm) {
+	const out = new Uint8Array(pcm.length);
+	for (let i = 0; i < pcm.length; i++) out[i] = mulawEncodeSample(pcm[i]);
+	return out;
+}
+/** Duration in milliseconds of a μ-law byte count at 8 kHz (1 byte = 125 µs). */
+function mulawBytesToMs(bytes) {
+	return bytes / MULAW_SAMPLE_RATE * 1e3;
+}
+/** Byte length of a base64 string without decoding it. */
+function base64ByteLength(base64) {
+	let padding = 0;
+	if (base64.endsWith("==")) padding = 2;
+	else if (base64.endsWith("=")) padding = 1;
+	return base64.length * 3 / 4 - padding;
+}
+//#endregion
+export { mulawBytesToMs as a, mulawToPcm16 as c, base64ByteLength as i, pcm16ToMulaw as l, MULAW_SAMPLE_RATE as n, mulawDecodeSample as o, MULAW_SILENCE_BYTE as r, mulawEncodeSample as s, MULAW_FRAME_BYTES_20MS as t };

package/dist/mulaw-DLUObjdP.cjs

@@ -0,0 +1,117 @@
+//#region src/audio/mulaw.ts
+/**
+* G.711 μ-law codec (the format Twilio Media Streams speak: 8 kHz, 1 byte/sample).
+*
+* Decode uses a 256-entry lookup table; encode uses the canonical Sun g711.c
+* algorithm (bias 0x84, clip 32635). μ-law silence is 0xFF.
+*/
+/** One byte of μ-law silence (encodes linear 0). */
+const MULAW_SILENCE_BYTE = 255;
+/** Samples per second on the Twilio media stream. μ-law is 1 byte per sample. */
+const MULAW_SAMPLE_RATE = 8e3;
+/** Bytes per 20 ms Twilio media frame at 8 kHz μ-law. */
+const MULAW_FRAME_BYTES_20MS = 160;
+function decodeSampleUncached(byte) {
+	const u = ~byte & 255;
+	let t = ((u & 15) << 3) + 132;
+	t <<= (u & 112) >> 4;
+	return u & 128 ? 132 - t : t - 132;
+}
+const DECODE_TABLE = /* @__PURE__ */ new Int16Array(256);
+for (let i = 0; i < 256; i++) DECODE_TABLE[i] = decodeSampleUncached(i);
+const BIAS = 132;
+const CLIP = 32635;
+/** Decode a single μ-law byte to a linear PCM16 sample. */
+function mulawDecodeSample(byte) {
+	return DECODE_TABLE[byte & 255];
+}
+/** Encode a single linear PCM16 sample to a μ-law byte. */
+function mulawEncodeSample(sample) {
+	let pcm = sample;
+	const sign = pcm < 0 ? 128 : 0;
+	if (sign) pcm = -pcm;
+	if (pcm > CLIP) pcm = CLIP;
+	pcm += BIAS;
+	let exponent = 7;
+	for (let mask = 16384; (pcm & mask) === 0 && exponent > 0; mask >>= 1) exponent--;
+	const mantissa = pcm >> exponent + 3 & 15;
+	return ~(sign | exponent << 4 | mantissa) & 255;
+}
+/** Decode a μ-law buffer to linear PCM16 samples. */
+function mulawToPcm16(mulaw) {
+	const out = new Int16Array(mulaw.length);
+	for (let i = 0; i < mulaw.length; i++) out[i] = DECODE_TABLE[mulaw[i]];
+	return out;
+}
+/** Encode linear PCM16 samples to a μ-law buffer. */
+function pcm16ToMulaw(pcm) {
+	const out = new Uint8Array(pcm.length);
+	for (let i = 0; i < pcm.length; i++) out[i] = mulawEncodeSample(pcm[i]);
+	return out;
+}
+/** Duration in milliseconds of a μ-law byte count at 8 kHz (1 byte = 125 µs). */
+function mulawBytesToMs(bytes) {
+	return bytes / MULAW_SAMPLE_RATE * 1e3;
+}
+/** Byte length of a base64 string without decoding it. */
+function base64ByteLength(base64) {
+	let padding = 0;
+	if (base64.endsWith("==")) padding = 2;
+	else if (base64.endsWith("=")) padding = 1;
+	return base64.length * 3 / 4 - padding;
+}
+//#endregion
+Object.defineProperty(exports, "MULAW_FRAME_BYTES_20MS", {
+	enumerable: true,
+	get: function() {
+		return MULAW_FRAME_BYTES_20MS;
+	}
+});
+Object.defineProperty(exports, "MULAW_SAMPLE_RATE", {
+	enumerable: true,
+	get: function() {
+		return MULAW_SAMPLE_RATE;
+	}
+});
+Object.defineProperty(exports, "MULAW_SILENCE_BYTE", {
+	enumerable: true,
+	get: function() {
+		return MULAW_SILENCE_BYTE;
+	}
+});
+Object.defineProperty(exports, "base64ByteLength", {
+	enumerable: true,
+	get: function() {
+		return base64ByteLength;
+	}
+});
+Object.defineProperty(exports, "mulawBytesToMs", {
+	enumerable: true,
+	get: function() {
+		return mulawBytesToMs;
+	}
+});
+Object.defineProperty(exports, "mulawDecodeSample", {
+	enumerable: true,
+	get: function() {
+		return mulawDecodeSample;
+	}
+});
+Object.defineProperty(exports, "mulawEncodeSample", {
+	enumerable: true,
+	get: function() {
+		return mulawEncodeSample;
+	}
+});
+Object.defineProperty(exports, "mulawToPcm16", {
+	enumerable: true,
+	get: function() {
+		return mulawToPcm16;
+	}
+});
+Object.defineProperty(exports, "pcm16ToMulaw", {
+	enumerable: true,
+	get: function() {
+		return pcm16ToMulaw;
+	}
+});

package/dist/openai.cjs

@@ -0,0 +1,61 @@
+Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
+const require_session_config = require("./session-config-c8sOw1XL.cjs");
+const require_env = require("./env-DSnGaERV.cjs");
+const require_OpenAICompatibleProvider = require("./OpenAICompatibleProvider-NS4cKVQj.cjs");
+//#region src/openai.ts
+/** OpenAI Realtime provider (GA API) — `realtime-voice-agents/openai`. */
+/** Env vars checked (in order) when no explicit apiKey is passed. */
+const OPENAI_KEY_ENV_VARS = [
+	"OPENAI_API_KEY",
+	"OPENAI_KEY",
+	"OPEN_AI_API_KEY"
+];
+const OPENAI_DEFAULT_MODEL = "gpt-realtime";
+const OPENAI_DEFAULT_VOICE = "marin";
+/**
+* Voices supported by realtime models. OpenAI recommends marin/cedar (the
+* realtime-native ones — audibly better once 8 kHz telephony takes its cut).
+* The legacy TTS voices (fable, onyx, nova) are NOT supported and fail in
+* unpleasant ways instead of erroring cleanly.
+*/
+const OPENAI_REALTIME_VOICES = [
+	"marin",
+	"cedar",
+	"alloy",
+	"ash",
+	"ballad",
+	"coral",
+	"echo",
+	"sage",
+	"shimmer",
+	"verse"
+];
+/** Create a provider factory for the bridge (`provider: openaiRealtime({...})`). */
+function openaiRealtime(options = {}) {
+	const apiKey = require_env.resolveApiKey(options.apiKey, OPENAI_KEY_ENV_VARS);
+	if (!apiKey) throw new Error(`openaiRealtime: apiKey missing (pass apiKey or set one of ${OPENAI_KEY_ENV_VARS.join("/")})`);
+	const config = {
+		apiKey,
+		model: options.model ?? "gpt-realtime",
+		voice: options.voice ?? "marin",
+		baseUrl: options.baseUrl,
+		defaultVad: options.vad,
+		defaultTranscription: options.transcription,
+		headers: options.headers,
+		extraSessionOptions: options.sessionOptions,
+		connectTimeoutMs: options.connectTimeoutMs,
+		providerName: "openai"
+	};
+	return ({ logger }) => {
+		return new require_OpenAICompatibleProvider.OpenAICompatibleProvider(config, logger);
+	};
+}
+//#endregion
+exports.OPENAI_DEFAULT_MODEL = OPENAI_DEFAULT_MODEL;
+exports.OPENAI_DEFAULT_VOICE = OPENAI_DEFAULT_VOICE;
+exports.OPENAI_KEY_ENV_VARS = OPENAI_KEY_ENV_VARS;
+exports.OPENAI_REALTIME_VOICES = OPENAI_REALTIME_VOICES;
+exports.OpenAICompatibleProvider = require_OpenAICompatibleProvider.OpenAICompatibleProvider;
+exports.buildSessionUpdate = require_session_config.buildSessionUpdate;
+exports.buildTurnDetection = require_session_config.buildTurnDetection;
+exports.openaiRealtime = openaiRealtime;

package/dist/openai.d.cts

@@ -0,0 +1,121 @@
+import { c as VadConfig, h as Logger, i as ProviderSessionInit, m as ProviderCapabilities, n as ProviderFactory, o as SendTextOptions, s as SendToolResultOptions, t as BaseRealtimeProvider } from "./BaseRealtimeProvider-BehPNT1r.cjs";
+//#region src/providers/openai-compatible/OpenAICompatibleProvider.d.ts
+interface OpenAICompatibleProviderConfig {
+  apiKey: string;
+  model: string;
+  /** WS endpoint; `?model=` is appended. Default OpenAI's realtime endpoint. */
+  baseUrl?: string;
+  voice?: string;
+  /** VAD used when the session init doesn't specify one. `null` disables. */
+  defaultVad?: VadConfig | null;
+  /** Transcription used when the session init doesn't specify one. */
+  defaultTranscription?: {
+    model?: string;
+    language?: string;
+  } | false;
+  headers?: Record<string, string>;
+  /** Provider-native session fields, deep-merged into session.update last. */
+  extraSessionOptions?: Record<string, unknown>;
+  connectTimeoutMs?: number;
+  capabilityOverrides?: Partial<ProviderCapabilities>;
+  /** Provider display name for logs/events (e.g. 'openai', 'xai'). */
+  providerName?: string;
+  /**
+   * Session payload builder override for OpenAI-compatible services whose
+   * session shape diverges (xAI puts voice/turn_detection at the session
+   * root). Defaults to the OpenAI GA builder.
+   */
+  buildSession?: (init: ProviderSessionInit, config: {
+    defaultVoice?: string;
+    extraSessionOptions?: Record<string, unknown>;
+  }) => Record<string, unknown>;
+}
+declare class OpenAICompatibleProvider extends BaseRealtimeProvider {
+  readonly name: string;
+  readonly capabilities: ProviderCapabilities;
+  private readonly config;
+  private readonly logger;
+  private ws;
+  private sessionInit;
+  private ready;
+  private intentionalClose;
+  private currentResponseId;
+  /**
+   * response.create serialization. The GA API rejects a `response.create`
+   * issued while another response is in flight
+   * (`conversation_already_has_active_response`), so creates requested
+   * mid-response wait in a single pending slot and fire on `response.done`.
+   * Later requests coalesce into that slot — one response reads the whole
+   * conversation state, so only the instructions payload is worth keeping.
+   */
+  private responseActive;
+  private pendingCreate;
+  private lastCreateSent;
+  constructor(config: OpenAICompatibleProviderConfig, logger?: Logger);
+  get isConnected(): boolean;
+  connect(init: ProviderSessionInit): Promise<void>;
+  close(): Promise<void>;
+  sendAudio(base64Mulaw: string): void;
+  sendText(text: string, options?: SendTextOptions): void;
+  sendToolResult(callId: string, output: unknown, options?: SendToolResultOptions): void;
+  createResponse(options?: {
+    instructions?: string;
+  }): void;
+  private sendCreate;
+  updateSession(patch: Partial<ProviderSessionInit>): Promise<void>;
+  private buildSessionPayload;
+  cancelResponse(): void;
+  truncatePlayback(itemId: string, audioEndMs: number): void;
+  private handleEvent;
+  private parseEvent;
+  private send;
+}
+//#endregion
+//#region src/providers/openai-compatible/session-config.d.ts
+declare function buildTurnDetection(vad: VadConfig | null | undefined): Record<string, unknown> | null;
+interface SessionPayloadConfig {
+  defaultVoice?: string;
+  /** Provider-native session fields, deep-merged last (escape hatch). */
+  extraSessionOptions?: Record<string, unknown>;
+}
+declare function buildSessionUpdate(init: ProviderSessionInit, config?: SessionPayloadConfig): Record<string, unknown>;
+//#endregion
+//#region src/openai.d.ts
+/** Env vars checked (in order) when no explicit apiKey is passed. */
+declare const OPENAI_KEY_ENV_VARS: readonly ["OPENAI_API_KEY", "OPENAI_KEY", "OPEN_AI_API_KEY"];
+declare const OPENAI_DEFAULT_MODEL = "gpt-realtime";
+declare const OPENAI_DEFAULT_VOICE = "marin";
+/**
+ * Voices supported by realtime models. OpenAI recommends marin/cedar (the
+ * realtime-native ones — audibly better once 8 kHz telephony takes its cut).
+ * The legacy TTS voices (fable, onyx, nova) are NOT supported and fail in
+ * unpleasant ways instead of erroring cleanly.
+ */
+declare const OPENAI_REALTIME_VOICES: readonly ["marin", "cedar", "alloy", "ash", "ballad", "coral", "echo", "sage", "shimmer", "verse"];
+type OpenAIRealtimeVoice = (typeof OPENAI_REALTIME_VOICES)[number];
+interface OpenAIRealtimeOptions {
+  /** Defaults to the first of OPENAI_API_KEY / OPENAI_KEY / OPEN_AI_API_KEY set in the env. */
+  apiKey?: string;
+  /** Realtime model id. Default `gpt-realtime`. */
+  model?: string;
+  /** Default voice when the Agent doesn't set one. Default `marin`. */
+  voice?: string;
+  /** Override the WS endpoint (proxies, gateways). */
+  baseUrl?: string;
+  /** Normalized VAD config; `null` disables server turn detection. */
+  vad?: VadConfig | null;
+  /** Caller-speech transcription; `false` disables it. */
+  transcription?: {
+    model?: string;
+    language?: string;
+  } | false;
+  /** Extra WS headers. */
+  headers?: Record<string, string>;
+  /** Provider-native session fields, deep-merged into session.update last. */
+  sessionOptions?: Record<string, unknown>;
+  connectTimeoutMs?: number;
+}
+/** Create a provider factory for the bridge (`provider: openaiRealtime({...})`). */
+declare function openaiRealtime(options?: OpenAIRealtimeOptions): ProviderFactory;
+//#endregion
+export { OPENAI_DEFAULT_MODEL, OPENAI_DEFAULT_VOICE, OPENAI_KEY_ENV_VARS, OPENAI_REALTIME_VOICES, OpenAICompatibleProvider, type OpenAICompatibleProviderConfig, OpenAIRealtimeOptions, OpenAIRealtimeVoice, buildSessionUpdate, buildTurnDetection, openaiRealtime };

package/dist/openai.d.mts

@@ -0,0 +1,121 @@
+import { c as VadConfig, h as Logger, i as ProviderSessionInit, m as ProviderCapabilities, n as ProviderFactory, o as SendTextOptions, s as SendToolResultOptions, t as BaseRealtimeProvider } from "./BaseRealtimeProvider-ClP8Wx1X.mjs";
+//#region src/providers/openai-compatible/OpenAICompatibleProvider.d.ts
+interface OpenAICompatibleProviderConfig {
+  apiKey: string;
+  model: string;
+  /** WS endpoint; `?model=` is appended. Default OpenAI's realtime endpoint. */
+  baseUrl?: string;
+  voice?: string;
+  /** VAD used when the session init doesn't specify one. `null` disables. */
+  defaultVad?: VadConfig | null;
+  /** Transcription used when the session init doesn't specify one. */
+  defaultTranscription?: {
+    model?: string;
+    language?: string;
+  } | false;
+  headers?: Record<string, string>;
+  /** Provider-native session fields, deep-merged into session.update last. */
+  extraSessionOptions?: Record<string, unknown>;
+  connectTimeoutMs?: number;
+  capabilityOverrides?: Partial<ProviderCapabilities>;
+  /** Provider display name for logs/events (e.g. 'openai', 'xai'). */
+  providerName?: string;
+  /**
+   * Session payload builder override for OpenAI-compatible services whose
+   * session shape diverges (xAI puts voice/turn_detection at the session
+   * root). Defaults to the OpenAI GA builder.
+   */
+  buildSession?: (init: ProviderSessionInit, config: {
+    defaultVoice?: string;
+    extraSessionOptions?: Record<string, unknown>;
+  }) => Record<string, unknown>;
+}
+declare class OpenAICompatibleProvider extends BaseRealtimeProvider {
+  readonly name: string;
+  readonly capabilities: ProviderCapabilities;
+  private readonly config;
+  private readonly logger;
+  private ws;
+  private sessionInit;
+  private ready;
+  private intentionalClose;
+  private currentResponseId;
+  /**
+   * response.create serialization. The GA API rejects a `response.create`
+   * issued while another response is in flight
+   * (`conversation_already_has_active_response`), so creates requested
+   * mid-response wait in a single pending slot and fire on `response.done`.
+   * Later requests coalesce into that slot — one response reads the whole
+   * conversation state, so only the instructions payload is worth keeping.
+   */
+  private responseActive;
+  private pendingCreate;
+  private lastCreateSent;
+  constructor(config: OpenAICompatibleProviderConfig, logger?: Logger);
+  get isConnected(): boolean;
+  connect(init: ProviderSessionInit): Promise<void>;
+  close(): Promise<void>;
+  sendAudio(base64Mulaw: string): void;
+  sendText(text: string, options?: SendTextOptions): void;
+  sendToolResult(callId: string, output: unknown, options?: SendToolResultOptions): void;
+  createResponse(options?: {
+    instructions?: string;
+  }): void;
+  private sendCreate;
+  updateSession(patch: Partial<ProviderSessionInit>): Promise<void>;
+  private buildSessionPayload;
+  cancelResponse(): void;
+  truncatePlayback(itemId: string, audioEndMs: number): void;
+  private handleEvent;
+  private parseEvent;
+  private send;
+}
+//#endregion
+//#region src/providers/openai-compatible/session-config.d.ts
+declare function buildTurnDetection(vad: VadConfig | null | undefined): Record<string, unknown> | null;
+interface SessionPayloadConfig {
+  defaultVoice?: string;
+  /** Provider-native session fields, deep-merged last (escape hatch). */
+  extraSessionOptions?: Record<string, unknown>;
+}
+declare function buildSessionUpdate(init: ProviderSessionInit, config?: SessionPayloadConfig): Record<string, unknown>;
+//#endregion
+//#region src/openai.d.ts
+/** Env vars checked (in order) when no explicit apiKey is passed. */
+declare const OPENAI_KEY_ENV_VARS: readonly ["OPENAI_API_KEY", "OPENAI_KEY", "OPEN_AI_API_KEY"];
+declare const OPENAI_DEFAULT_MODEL = "gpt-realtime";
+declare const OPENAI_DEFAULT_VOICE = "marin";
+/**
+ * Voices supported by realtime models. OpenAI recommends marin/cedar (the
+ * realtime-native ones — audibly better once 8 kHz telephony takes its cut).
+ * The legacy TTS voices (fable, onyx, nova) are NOT supported and fail in
+ * unpleasant ways instead of erroring cleanly.
+ */
+declare const OPENAI_REALTIME_VOICES: readonly ["marin", "cedar", "alloy", "ash", "ballad", "coral", "echo", "sage", "shimmer", "verse"];
+type OpenAIRealtimeVoice = (typeof OPENAI_REALTIME_VOICES)[number];
+interface OpenAIRealtimeOptions {
+  /** Defaults to the first of OPENAI_API_KEY / OPENAI_KEY / OPEN_AI_API_KEY set in the env. */
+  apiKey?: string;
+  /** Realtime model id. Default `gpt-realtime`. */
+  model?: string;
+  /** Default voice when the Agent doesn't set one. Default `marin`. */
+  voice?: string;
+  /** Override the WS endpoint (proxies, gateways). */
+  baseUrl?: string;
+  /** Normalized VAD config; `null` disables server turn detection. */
+  vad?: VadConfig | null;
+  /** Caller-speech transcription; `false` disables it. */
+  transcription?: {
+    model?: string;
+    language?: string;
+  } | false;
+  /** Extra WS headers. */
+  headers?: Record<string, string>;
+  /** Provider-native session fields, deep-merged into session.update last. */
+  sessionOptions?: Record<string, unknown>;
+  connectTimeoutMs?: number;
+}
+/** Create a provider factory for the bridge (`provider: openaiRealtime({...})`). */
+declare function openaiRealtime(options?: OpenAIRealtimeOptions): ProviderFactory;
+//#endregion
+export { OPENAI_DEFAULT_MODEL, OPENAI_DEFAULT_VOICE, OPENAI_KEY_ENV_VARS, OPENAI_REALTIME_VOICES, OpenAICompatibleProvider, type OpenAICompatibleProviderConfig, OpenAIRealtimeOptions, OpenAIRealtimeVoice, buildSessionUpdate, buildTurnDetection, openaiRealtime };

package/dist/openai.mjs

@@ -0,0 +1,53 @@
+import { n as buildTurnDetection, t as buildSessionUpdate } from "./session-config-BVLl7-ha.mjs";
+import { t as resolveApiKey } from "./env-DUwUWTsg.mjs";
+import { t as OpenAICompatibleProvider } from "./OpenAICompatibleProvider-Bdtl-UXH.mjs";
+//#region src/openai.ts
+/** OpenAI Realtime provider (GA API) — `realtime-voice-agents/openai`. */
+/** Env vars checked (in order) when no explicit apiKey is passed. */
+const OPENAI_KEY_ENV_VARS = [
+	"OPENAI_API_KEY",
+	"OPENAI_KEY",
+	"OPEN_AI_API_KEY"
+];
+const OPENAI_DEFAULT_MODEL = "gpt-realtime";
+const OPENAI_DEFAULT_VOICE = "marin";
+/**
+* Voices supported by realtime models. OpenAI recommends marin/cedar (the
+* realtime-native ones — audibly better once 8 kHz telephony takes its cut).
+* The legacy TTS voices (fable, onyx, nova) are NOT supported and fail in
+* unpleasant ways instead of erroring cleanly.
+*/
+const OPENAI_REALTIME_VOICES = [
+	"marin",
+	"cedar",
+	"alloy",
+	"ash",
+	"ballad",
+	"coral",
+	"echo",
+	"sage",
+	"shimmer",
+	"verse"
+];
+/** Create a provider factory for the bridge (`provider: openaiRealtime({...})`). */
+function openaiRealtime(options = {}) {
+	const apiKey = resolveApiKey(options.apiKey, OPENAI_KEY_ENV_VARS);
+	if (!apiKey) throw new Error(`openaiRealtime: apiKey missing (pass apiKey or set one of ${OPENAI_KEY_ENV_VARS.join("/")})`);
+	const config = {
+		apiKey,
+		model: options.model ?? "gpt-realtime",
+		voice: options.voice ?? "marin",
+		baseUrl: options.baseUrl,
+		defaultVad: options.vad,
+		defaultTranscription: options.transcription,
+		headers: options.headers,
+		extraSessionOptions: options.sessionOptions,
+		connectTimeoutMs: options.connectTimeoutMs,
+		providerName: "openai"
+	};
+	return ({ logger }) => {
+		return new OpenAICompatibleProvider(config, logger);
+	};
+}
+//#endregion
+export { OPENAI_DEFAULT_MODEL, OPENAI_DEFAULT_VOICE, OPENAI_KEY_ENV_VARS, OPENAI_REALTIME_VOICES, OpenAICompatibleProvider, buildSessionUpdate, buildTurnDetection, openaiRealtime };

package/dist/presets-Bf75YXs5.d.cts

@@ -0,0 +1,24 @@
+//#region src/audio/background/presets.d.ts
+/**
+ * Background-audio sources. Presets resolve to bundled μ-law loop assets;
+ * custom sources accept a μ-law buffer or a file path to one.
+ * (The playback engine lands with the background-audio milestone; the types
+ * live here so tool definitions can reference them.)
+ */
+type BackgroundAudioPreset = 'elevator-jazz' | 'lofi' | 'keyboard-typing' | 'thinking-hum' | 'ringing';
+type BackgroundAudioSpec = BackgroundAudioPreset | {
+  /** Raw 8 kHz μ-law audio, or an absolute path to a `.ulaw` file. */
+  custom: Buffer | string;
+};
+interface BackgroundAudioOptions {
+  /** Linear gain 0..1 applied to the loop. Default 0.4. */
+  volume?: number;
+  fadeInMs?: number;
+  fadeOutMs?: number;
+  /** Delay before audio actually starts — fast tools never trigger it. */
+  startDelayMs?: number;
+  /** Hard stop even if never released. Default 60000. */
+  maxDurationMs?: number;
+}
+//#endregion
+export { BackgroundAudioPreset as n, BackgroundAudioSpec as r, BackgroundAudioOptions as t };

package/dist/presets-Bf75YXs5.d.mts

@@ -0,0 +1,24 @@
+//#region src/audio/background/presets.d.ts
+/**
+ * Background-audio sources. Presets resolve to bundled μ-law loop assets;
+ * custom sources accept a μ-law buffer or a file path to one.
+ * (The playback engine lands with the background-audio milestone; the types
+ * live here so tool definitions can reference them.)
+ */
+type BackgroundAudioPreset = 'elevator-jazz' | 'lofi' | 'keyboard-typing' | 'thinking-hum' | 'ringing';
+type BackgroundAudioSpec = BackgroundAudioPreset | {
+  /** Raw 8 kHz μ-law audio, or an absolute path to a `.ulaw` file. */
+  custom: Buffer | string;
+};
+interface BackgroundAudioOptions {
+  /** Linear gain 0..1 applied to the loop. Default 0.4. */
+  volume?: number;
+  fadeInMs?: number;
+  fadeOutMs?: number;
+  /** Delay before audio actually starts — fast tools never trigger it. */
+  startDelayMs?: number;
+  /** Hard stop even if never released. Default 60000. */
+  maxDurationMs?: number;
+}
+//#endregion
+export { BackgroundAudioPreset as n, BackgroundAudioSpec as r, BackgroundAudioOptions as t };