@livekit/agents 1.0.34 → 1.0.36-dev.0

package/dist/inference/interruption/http_transport.js

@@ -0,0 +1,33 @@
+import { ofetch } from "ofetch";
+async function predictHTTP(data, predictOptions, options) {
+  const createdAt = performance.now();
+  const url = new URL(`/bargein`, options.baseUrl);
+  url.searchParams.append("threshold", predictOptions.threshold.toString());
+  url.searchParams.append("min_frames", predictOptions.minFrames.toFixed());
+  url.searchParams.append("created_at", createdAt.toFixed());
+  const { created_at, is_bargein, probabilities } = await ofetch(
+    url.toString(),
+    {
+      retry: 1,
+      retryDelay: 100,
+      headers: {
+        "Content-Type": "application/octet-stream",
+        Authorization: `Bearer ${options.token}`
+      },
+      signal: options.signal,
+      timeout: options.timeout,
+      method: "POST",
+      body: data
+    }
+  );
+  return {
+    createdAt: created_at,
+    isBargein: is_bargein,
+    probabilities: new Float32Array(probabilities),
+    predictionDuration: (performance.now() - createdAt) / 1e9
+  };
+}
+export {
+  predictHTTP
+};
+//# sourceMappingURL=http_transport.js.map

package/dist/inference/interruption/http_transport.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/inference/interruption/http_transport.ts"],"sourcesContent":["import { ofetch } from 'ofetch';\n\nexport interface PostOptions {\n  baseUrl: string;\n  token: string;\n  signal?: AbortSignal;\n  timeout?: number;\n}\n\nexport interface PredictOptions {\n  threshold: number;\n  minFrames: number;\n}\n\nexport interface PredictEndpointResponse {\n  created_at: number;\n  is_bargein: boolean;\n  probabilities: number[];\n}\n\nexport interface PredictResponse {\n  createdAt: number;\n  isBargein: boolean;\n  probabilities: Float32Array;\n  predictionDuration: number;\n}\n\nexport async function predictHTTP(\n  data: Int16Array,\n  predictOptions: PredictOptions,\n  options: PostOptions,\n): Promise<PredictResponse> {\n  const createdAt = performance.now();\n  const url = new URL(`/bargein`, options.baseUrl);\n  url.searchParams.append('threshold', predictOptions.threshold.toString());\n  url.searchParams.append('min_frames', predictOptions.minFrames.toFixed());\n  url.searchParams.append('created_at', createdAt.toFixed());\n\n  const { created_at, is_bargein, probabilities } = await ofetch<PredictEndpointResponse>(\n    url.toString(),\n    {\n      retry: 1,\n      retryDelay: 100,\n      headers: {\n        'Content-Type': 'application/octet-stream',\n        Authorization: `Bearer ${options.token}`,\n      },\n      signal: options.signal,\n      timeout: options.timeout,\n      method: 'POST',\n      body: data,\n    },\n  );\n\n  return {\n    createdAt: created_at,\n    isBargein: is_bargein,\n    probabilities: new Float32Array(probabilities),\n    predictionDuration: (performance.now() - createdAt) / 1e9,\n  };\n}\n"],"mappings":"AAAA,SAAS,cAAc;AA2BvB,eAAsB,YACpB,MACA,gBACA,SAC0B;AAC1B,QAAM,YAAY,YAAY,IAAI;AAClC,QAAM,MAAM,IAAI,IAAI,YAAY,QAAQ,OAAO;AAC/C,MAAI,aAAa,OAAO,aAAa,eAAe,UAAU,SAAS,CAAC;AACxE,MAAI,aAAa,OAAO,cAAc,eAAe,UAAU,QAAQ,CAAC;AACxE,MAAI,aAAa,OAAO,cAAc,UAAU,QAAQ,CAAC;AAEzD,QAAM,EAAE,YAAY,YAAY,cAAc,IAAI,MAAM;AAAA,IACtD,IAAI,SAAS;AAAA,IACb;AAAA,MACE,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,SAAS;AAAA,QACP,gBAAgB;AAAA,QAChB,eAAe,UAAU,QAAQ,KAAK;AAAA,MACxC;AAAA,MACA,QAAQ,QAAQ;AAAA,MAChB,SAAS,QAAQ;AAAA,MACjB,QAAQ;AAAA,MACR,MAAM;AAAA,IACR;AAAA,EACF;AAEA,SAAO;AAAA,IACL,WAAW;AAAA,IACX,WAAW;AAAA,IACX,eAAe,IAAI,aAAa,aAAa;AAAA,IAC7C,qBAAqB,YAAY,IAAI,IAAI,aAAa;AAAA,EACxD;AACF;","names":[]}

package/dist/inference/interruption/index.cjs

@@ -0,0 +1,34 @@
+"use strict";
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __reExport = (target, mod, secondTarget) => (__copyProps(target, mod, "default"), secondTarget && __copyProps(secondTarget, mod, "default"));
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+var interruption_exports = {};
+__export(interruption_exports, {
+  InterruptionStreamSentinel: () => import_InterruptionStream.InterruptionStreamSentinel
+});
+module.exports = __toCommonJS(interruption_exports);
+__reExport(interruption_exports, require("./AdaptiveInterruptionDetector.cjs"), module.exports);
+__reExport(interruption_exports, require("./interruption.cjs"), module.exports);
+var import_InterruptionStream = require("./InterruptionStream.cjs");
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  InterruptionStreamSentinel,
+  ...require("./AdaptiveInterruptionDetector.cjs"),
+  ...require("./interruption.cjs")
+});
+//# sourceMappingURL=index.cjs.map

package/dist/inference/interruption/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/inference/interruption/index.ts"],"sourcesContent":["export * from './AdaptiveInterruptionDetector.js';\nexport * from './interruption.js';\nexport { InterruptionStreamSentinel } from './InterruptionStream.js';\nexport type { InterruptionSentinel } from './InterruptionStream.js';\n"],"mappings":";;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,iCAAc,8CAAd;AACA,iCAAc,8BADd;AAEA,gCAA2C;","names":[]}

package/dist/inference/interruption/index.d.cts

@@ -0,0 +1,5 @@
+export * from './AdaptiveInterruptionDetector.js';
+export * from './interruption.js';
+export { InterruptionStreamSentinel } from './InterruptionStream.js';
+export type { InterruptionSentinel } from './InterruptionStream.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/inference/interruption/index.d.ts

@@ -0,0 +1,5 @@
+export * from './AdaptiveInterruptionDetector.js';
+export * from './interruption.js';
+export { InterruptionStreamSentinel } from './InterruptionStream.js';
+export type { InterruptionSentinel } from './InterruptionStream.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/inference/interruption/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/inference/interruption/index.ts"],"names":[],"mappings":"AAAA,cAAc,mCAAmC,CAAC;AAClD,cAAc,mBAAmB,CAAC;AAClC,OAAO,EAAE,0BAA0B,EAAE,MAAM,yBAAyB,CAAC;AACrE,YAAY,EAAE,oBAAoB,EAAE,MAAM,yBAAyB,CAAC"}

package/dist/inference/interruption/index.js

@@ -0,0 +1,7 @@
+export * from "./AdaptiveInterruptionDetector.js";
+export * from "./interruption.js";
+import { InterruptionStreamSentinel } from "./InterruptionStream.js";
+export {
+  InterruptionStreamSentinel
+};
+//# sourceMappingURL=index.js.map

package/dist/inference/interruption/index.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/inference/interruption/index.ts"],"sourcesContent":["export * from './AdaptiveInterruptionDetector.js';\nexport * from './interruption.js';\nexport { InterruptionStreamSentinel } from './InterruptionStream.js';\nexport type { InterruptionSentinel } from './InterruptionStream.js';\n"],"mappings":"AAAA,cAAc;AACd,cAAc;AACd,SAAS,kCAAkC;","names":[]}

package/dist/inference/interruption/interruption.cjs

@@ -0,0 +1,85 @@
+"use strict";
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+var interruption_exports = {};
+__export(interruption_exports, {
+  InterruptionCacheEntry: () => InterruptionCacheEntry,
+  InterruptionDetectionError: () => InterruptionDetectionError,
+  InterruptionEventType: () => InterruptionEventType
+});
+module.exports = __toCommonJS(interruption_exports);
+var import_utils = require("../utils.cjs");
+var import_defaults = require("./defaults.cjs");
+var InterruptionEventType = /* @__PURE__ */ ((InterruptionEventType2) => {
+  InterruptionEventType2["INTERRUPTION"] = "interruption";
+  InterruptionEventType2["OVERLAP_SPEECH_ENDED"] = "overlap_speech_ended";
+  return InterruptionEventType2;
+})(InterruptionEventType || {});
+class InterruptionDetectionError extends Error {
+  type = "InterruptionDetectionError";
+  timestamp;
+  label;
+  recoverable;
+  constructor(message, timestamp, label, recoverable) {
+    super(message);
+    this.name = "InterruptionDetectionError";
+    this.timestamp = timestamp;
+    this.label = label;
+    this.recoverable = recoverable;
+  }
+  toString() {
+    return `${this.name}: ${this.message} (label=${this.label}, timestamp=${this.timestamp}, recoverable=${this.recoverable})`;
+  }
+}
+function estimateProbability(probabilities, windowSize = import_defaults.MIN_INTERRUPTION_DURATION) {
+  const minWindow = Math.ceil(windowSize / 0.025);
+  if (probabilities.length < minWindow) {
+    return 0;
+  }
+  return (0, import_utils.slidingWindowMinMax)(probabilities, windowSize);
+}
+class InterruptionCacheEntry {
+  createdAt;
+  totalDuration;
+  predictionDuration;
+  detectionDelay;
+  speechInput;
+  probabilities;
+  isInterruption;
+  probability;
+  constructor(params) {
+    this.createdAt = params.createdAt;
+    this.totalDuration = params.totalDuration ?? 0;
+    this.predictionDuration = params.predictionDuration ?? 0;
+    this.detectionDelay = params.detectionDelay ?? 0;
+    this.speechInput = params.speechInput;
+    this.probabilities = params.probabilities;
+    this.isInterruption = params.isInterruption;
+    this.probability = this.probabilities ? estimateProbability(this.probabilities) : 0;
+  }
+  static default() {
+    return new InterruptionCacheEntry({ createdAt: 0 });
+  }
+}
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  InterruptionCacheEntry,
+  InterruptionDetectionError,
+  InterruptionEventType
+});
+//# sourceMappingURL=interruption.cjs.map

package/dist/inference/interruption/interruption.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/inference/interruption/interruption.ts"],"sourcesContent":["import { slidingWindowMinMax } from '../utils.js';\nimport { MIN_INTERRUPTION_DURATION } from './defaults.js';\n\nexport enum InterruptionEventType {\n  INTERRUPTION = 'interruption',\n  OVERLAP_SPEECH_ENDED = 'overlap_speech_ended',\n}\nexport interface InterruptionEvent {\n  type: InterruptionEventType;\n  timestamp: number;\n  isInterruption: boolean;\n  totalDuration: number;\n  predictionDuration: number;\n  detectionDelay: number;\n  overlapSpeechStartedAt?: number;\n  speechInput?: Int16Array;\n  probabilities?: Float32Array;\n  probability: number;\n}\n\nexport class InterruptionDetectionError extends Error {\n  readonly type = 'InterruptionDetectionError';\n\n  readonly timestamp: number;\n  readonly label: string;\n  readonly recoverable: boolean;\n\n  constructor(message: string, timestamp: number, label: string, recoverable: boolean) {\n    super(message);\n    this.name = 'InterruptionDetectionError';\n    this.timestamp = timestamp;\n    this.label = label;\n    this.recoverable = recoverable;\n  }\n\n  toString(): string {\n    return `${this.name}: ${this.message} (label=${this.label}, timestamp=${this.timestamp}, recoverable=${this.recoverable})`;\n  }\n}\n\nfunction estimateProbability(\n  probabilities: Float32Array,\n  windowSize: number = MIN_INTERRUPTION_DURATION,\n): number {\n  const minWindow = Math.ceil(windowSize / 0.025); // 25ms per frame\n  if (probabilities.length < minWindow) {\n    return 0;\n  }\n\n  return slidingWindowMinMax(probabilities, windowSize);\n}\n\n/**\n * Typed cache entry for interruption inference results.\n */\nexport class InterruptionCacheEntry {\n  readonly createdAt: number;\n  readonly totalDuration: number;\n  readonly predictionDuration: number;\n  readonly detectionDelay: number;\n  readonly speechInput?: Int16Array;\n  readonly probabilities?: Float32Array;\n  readonly isInterruption?: boolean;\n  readonly probability: number;\n\n  constructor(params: {\n    createdAt: number;\n    speechInput?: Int16Array;\n    totalDuration?: number;\n    predictionDuration?: number;\n    detectionDelay?: number;\n    probabilities?: Float32Array;\n    isInterruption?: boolean;\n  }) {\n    this.createdAt = params.createdAt;\n    this.totalDuration = params.totalDuration ?? 0;\n    this.predictionDuration = params.predictionDuration ?? 0;\n    this.detectionDelay = params.detectionDelay ?? 0;\n    this.speechInput = params.speechInput;\n    this.probabilities = params.probabilities;\n    this.isInterruption = params.isInterruption;\n    this.probability = this.probabilities ? estimateProbability(this.probabilities) : 0;\n  }\n\n  static default(): InterruptionCacheEntry {\n    return new InterruptionCacheEntry({ createdAt: 0 });\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,mBAAoC;AACpC,sBAA0C;AAEnC,IAAK,wBAAL,kBAAKA,2BAAL;AACL,EAAAA,uBAAA,kBAAe;AACf,EAAAA,uBAAA,0BAAuB;AAFb,SAAAA;AAAA,GAAA;AAiBL,MAAM,mCAAmC,MAAM;AAAA,EAC3C,OAAO;AAAA,EAEP;AAAA,EACA;AAAA,EACA;AAAA,EAET,YAAY,SAAiB,WAAmB,OAAe,aAAsB;AACnF,UAAM,OAAO;AACb,SAAK,OAAO;AACZ,SAAK,YAAY;AACjB,SAAK,QAAQ;AACb,SAAK,cAAc;AAAA,EACrB;AAAA,EAEA,WAAmB;AACjB,WAAO,GAAG,KAAK,IAAI,KAAK,KAAK,OAAO,WAAW,KAAK,KAAK,eAAe,KAAK,SAAS,iBAAiB,KAAK,WAAW;AAAA,EACzH;AACF;AAEA,SAAS,oBACP,eACA,aAAqB,2CACb;AACR,QAAM,YAAY,KAAK,KAAK,aAAa,KAAK;AAC9C,MAAI,cAAc,SAAS,WAAW;AACpC,WAAO;AAAA,EACT;AAEA,aAAO,kCAAoB,eAAe,UAAU;AACtD;AAKO,MAAM,uBAAuB;AAAA,EACzB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAET,YAAY,QAQT;AACD,SAAK,YAAY,OAAO;AACxB,SAAK,gBAAgB,OAAO,iBAAiB;AAC7C,SAAK,qBAAqB,OAAO,sBAAsB;AACvD,SAAK,iBAAiB,OAAO,kBAAkB;AAC/C,SAAK,cAAc,OAAO;AAC1B,SAAK,gBAAgB,OAAO;AAC5B,SAAK,iBAAiB,OAAO;AAC7B,SAAK,cAAc,KAAK,gBAAgB,oBAAoB,KAAK,aAAa,IAAI;AAAA,EACpF;AAAA,EAEA,OAAO,UAAkC;AACvC,WAAO,IAAI,uBAAuB,EAAE,WAAW,EAAE,CAAC;AAAA,EACpD;AACF;","names":["InterruptionEventType"]}

package/dist/inference/interruption/interruption.d.cts

@@ -0,0 +1,48 @@
+export declare enum InterruptionEventType {
+    INTERRUPTION = "interruption",
+    OVERLAP_SPEECH_ENDED = "overlap_speech_ended"
+}
+export interface InterruptionEvent {
+    type: InterruptionEventType;
+    timestamp: number;
+    isInterruption: boolean;
+    totalDuration: number;
+    predictionDuration: number;
+    detectionDelay: number;
+    overlapSpeechStartedAt?: number;
+    speechInput?: Int16Array;
+    probabilities?: Float32Array;
+    probability: number;
+}
+export declare class InterruptionDetectionError extends Error {
+    readonly type = "InterruptionDetectionError";
+    readonly timestamp: number;
+    readonly label: string;
+    readonly recoverable: boolean;
+    constructor(message: string, timestamp: number, label: string, recoverable: boolean);
+    toString(): string;
+}
+/**
+ * Typed cache entry for interruption inference results.
+ */
+export declare class InterruptionCacheEntry {
+    readonly createdAt: number;
+    readonly totalDuration: number;
+    readonly predictionDuration: number;
+    readonly detectionDelay: number;
+    readonly speechInput?: Int16Array;
+    readonly probabilities?: Float32Array;
+    readonly isInterruption?: boolean;
+    readonly probability: number;
+    constructor(params: {
+        createdAt: number;
+        speechInput?: Int16Array;
+        totalDuration?: number;
+        predictionDuration?: number;
+        detectionDelay?: number;
+        probabilities?: Float32Array;
+        isInterruption?: boolean;
+    });
+    static default(): InterruptionCacheEntry;
+}
+//# sourceMappingURL=interruption.d.ts.map

package/dist/inference/interruption/interruption.d.ts

@@ -0,0 +1,48 @@
+export declare enum InterruptionEventType {
+    INTERRUPTION = "interruption",
+    OVERLAP_SPEECH_ENDED = "overlap_speech_ended"
+}
+export interface InterruptionEvent {
+    type: InterruptionEventType;
+    timestamp: number;
+    isInterruption: boolean;
+    totalDuration: number;
+    predictionDuration: number;
+    detectionDelay: number;
+    overlapSpeechStartedAt?: number;
+    speechInput?: Int16Array;
+    probabilities?: Float32Array;
+    probability: number;
+}
+export declare class InterruptionDetectionError extends Error {
+    readonly type = "InterruptionDetectionError";
+    readonly timestamp: number;
+    readonly label: string;
+    readonly recoverable: boolean;
+    constructor(message: string, timestamp: number, label: string, recoverable: boolean);
+    toString(): string;
+}
+/**
+ * Typed cache entry for interruption inference results.
+ */
+export declare class InterruptionCacheEntry {
+    readonly createdAt: number;
+    readonly totalDuration: number;
+    readonly predictionDuration: number;
+    readonly detectionDelay: number;
+    readonly speechInput?: Int16Array;
+    readonly probabilities?: Float32Array;
+    readonly isInterruption?: boolean;
+    readonly probability: number;
+    constructor(params: {
+        createdAt: number;
+        speechInput?: Int16Array;
+        totalDuration?: number;
+        predictionDuration?: number;
+        detectionDelay?: number;
+        probabilities?: Float32Array;
+        isInterruption?: boolean;
+    });
+    static default(): InterruptionCacheEntry;
+}
+//# sourceMappingURL=interruption.d.ts.map

package/dist/inference/interruption/interruption.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"interruption.d.ts","sourceRoot":"","sources":["../../../src/inference/interruption/interruption.ts"],"names":[],"mappings":"AAGA,oBAAY,qBAAqB;IAC/B,YAAY,iBAAiB;IAC7B,oBAAoB,yBAAyB;CAC9C;AACD,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,qBAAqB,CAAC;IAC5B,SAAS,EAAE,MAAM,CAAC;IAClB,cAAc,EAAE,OAAO,CAAC;IACxB,aAAa,EAAE,MAAM,CAAC;IACtB,kBAAkB,EAAE,MAAM,CAAC;IAC3B,cAAc,EAAE,MAAM,CAAC;IACvB,sBAAsB,CAAC,EAAE,MAAM,CAAC;IAChC,WAAW,CAAC,EAAE,UAAU,CAAC;IACzB,aAAa,CAAC,EAAE,YAAY,CAAC;IAC7B,WAAW,EAAE,MAAM,CAAC;CACrB;AAED,qBAAa,0BAA2B,SAAQ,KAAK;IACnD,QAAQ,CAAC,IAAI,gCAAgC;IAE7C,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,WAAW,EAAE,OAAO,CAAC;gBAElB,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,WAAW,EAAE,OAAO;IAQnF,QAAQ,IAAI,MAAM;CAGnB;AAcD;;GAEG;AACH,qBAAa,sBAAsB;IACjC,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;IACpC,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,WAAW,CAAC,EAAE,UAAU,CAAC;IAClC,QAAQ,CAAC,aAAa,CAAC,EAAE,YAAY,CAAC;IACtC,QAAQ,CAAC,cAAc,CAAC,EAAE,OAAO,CAAC;IAClC,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;gBAEjB,MAAM,EAAE;QAClB,SAAS,EAAE,MAAM,CAAC;QAClB,WAAW,CAAC,EAAE,UAAU,CAAC;QACzB,aAAa,CAAC,EAAE,MAAM,CAAC;QACvB,kBAAkB,CAAC,EAAE,MAAM,CAAC;QAC5B,cAAc,CAAC,EAAE,MAAM,CAAC;QACxB,aAAa,CAAC,EAAE,YAAY,CAAC;QAC7B,cAAc,CAAC,EAAE,OAAO,CAAC;KAC1B;IAWD,MAAM,CAAC,OAAO,IAAI,sBAAsB;CAGzC"}

package/dist/inference/interruption/interruption.js

@@ -0,0 +1,59 @@
+import { slidingWindowMinMax } from "../utils.js";
+import { MIN_INTERRUPTION_DURATION } from "./defaults.js";
+var InterruptionEventType = /* @__PURE__ */ ((InterruptionEventType2) => {
+  InterruptionEventType2["INTERRUPTION"] = "interruption";
+  InterruptionEventType2["OVERLAP_SPEECH_ENDED"] = "overlap_speech_ended";
+  return InterruptionEventType2;
+})(InterruptionEventType || {});
+class InterruptionDetectionError extends Error {
+  type = "InterruptionDetectionError";
+  timestamp;
+  label;
+  recoverable;
+  constructor(message, timestamp, label, recoverable) {
+    super(message);
+    this.name = "InterruptionDetectionError";
+    this.timestamp = timestamp;
+    this.label = label;
+    this.recoverable = recoverable;
+  }
+  toString() {
+    return `${this.name}: ${this.message} (label=${this.label}, timestamp=${this.timestamp}, recoverable=${this.recoverable})`;
+  }
+}
+function estimateProbability(probabilities, windowSize = MIN_INTERRUPTION_DURATION) {
+  const minWindow = Math.ceil(windowSize / 0.025);
+  if (probabilities.length < minWindow) {
+    return 0;
+  }
+  return slidingWindowMinMax(probabilities, windowSize);
+}
+class InterruptionCacheEntry {
+  createdAt;
+  totalDuration;
+  predictionDuration;
+  detectionDelay;
+  speechInput;
+  probabilities;
+  isInterruption;
+  probability;
+  constructor(params) {
+    this.createdAt = params.createdAt;
+    this.totalDuration = params.totalDuration ?? 0;
+    this.predictionDuration = params.predictionDuration ?? 0;
+    this.detectionDelay = params.detectionDelay ?? 0;
+    this.speechInput = params.speechInput;
+    this.probabilities = params.probabilities;
+    this.isInterruption = params.isInterruption;
+    this.probability = this.probabilities ? estimateProbability(this.probabilities) : 0;
+  }
+  static default() {
+    return new InterruptionCacheEntry({ createdAt: 0 });
+  }
+}
+export {
+  InterruptionCacheEntry,
+  InterruptionDetectionError,
+  InterruptionEventType
+};
+//# sourceMappingURL=interruption.js.map

package/dist/inference/interruption/interruption.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../../src/inference/interruption/interruption.ts"],"sourcesContent":["import { slidingWindowMinMax } from '../utils.js';\nimport { MIN_INTERRUPTION_DURATION } from './defaults.js';\n\nexport enum InterruptionEventType {\n  INTERRUPTION = 'interruption',\n  OVERLAP_SPEECH_ENDED = 'overlap_speech_ended',\n}\nexport interface InterruptionEvent {\n  type: InterruptionEventType;\n  timestamp: number;\n  isInterruption: boolean;\n  totalDuration: number;\n  predictionDuration: number;\n  detectionDelay: number;\n  overlapSpeechStartedAt?: number;\n  speechInput?: Int16Array;\n  probabilities?: Float32Array;\n  probability: number;\n}\n\nexport class InterruptionDetectionError extends Error {\n  readonly type = 'InterruptionDetectionError';\n\n  readonly timestamp: number;\n  readonly label: string;\n  readonly recoverable: boolean;\n\n  constructor(message: string, timestamp: number, label: string, recoverable: boolean) {\n    super(message);\n    this.name = 'InterruptionDetectionError';\n    this.timestamp = timestamp;\n    this.label = label;\n    this.recoverable = recoverable;\n  }\n\n  toString(): string {\n    return `${this.name}: ${this.message} (label=${this.label}, timestamp=${this.timestamp}, recoverable=${this.recoverable})`;\n  }\n}\n\nfunction estimateProbability(\n  probabilities: Float32Array,\n  windowSize: number = MIN_INTERRUPTION_DURATION,\n): number {\n  const minWindow = Math.ceil(windowSize / 0.025); // 25ms per frame\n  if (probabilities.length < minWindow) {\n    return 0;\n  }\n\n  return slidingWindowMinMax(probabilities, windowSize);\n}\n\n/**\n * Typed cache entry for interruption inference results.\n */\nexport class InterruptionCacheEntry {\n  readonly createdAt: number;\n  readonly totalDuration: number;\n  readonly predictionDuration: number;\n  readonly detectionDelay: number;\n  readonly speechInput?: Int16Array;\n  readonly probabilities?: Float32Array;\n  readonly isInterruption?: boolean;\n  readonly probability: number;\n\n  constructor(params: {\n    createdAt: number;\n    speechInput?: Int16Array;\n    totalDuration?: number;\n    predictionDuration?: number;\n    detectionDelay?: number;\n    probabilities?: Float32Array;\n    isInterruption?: boolean;\n  }) {\n    this.createdAt = params.createdAt;\n    this.totalDuration = params.totalDuration ?? 0;\n    this.predictionDuration = params.predictionDuration ?? 0;\n    this.detectionDelay = params.detectionDelay ?? 0;\n    this.speechInput = params.speechInput;\n    this.probabilities = params.probabilities;\n    this.isInterruption = params.isInterruption;\n    this.probability = this.probabilities ? estimateProbability(this.probabilities) : 0;\n  }\n\n  static default(): InterruptionCacheEntry {\n    return new InterruptionCacheEntry({ createdAt: 0 });\n  }\n}\n"],"mappings":"AAAA,SAAS,2BAA2B;AACpC,SAAS,iCAAiC;AAEnC,IAAK,wBAAL,kBAAKA,2BAAL;AACL,EAAAA,uBAAA,kBAAe;AACf,EAAAA,uBAAA,0BAAuB;AAFb,SAAAA;AAAA,GAAA;AAiBL,MAAM,mCAAmC,MAAM;AAAA,EAC3C,OAAO;AAAA,EAEP;AAAA,EACA;AAAA,EACA;AAAA,EAET,YAAY,SAAiB,WAAmB,OAAe,aAAsB;AACnF,UAAM,OAAO;AACb,SAAK,OAAO;AACZ,SAAK,YAAY;AACjB,SAAK,QAAQ;AACb,SAAK,cAAc;AAAA,EACrB;AAAA,EAEA,WAAmB;AACjB,WAAO,GAAG,KAAK,IAAI,KAAK,KAAK,OAAO,WAAW,KAAK,KAAK,eAAe,KAAK,SAAS,iBAAiB,KAAK,WAAW;AAAA,EACzH;AACF;AAEA,SAAS,oBACP,eACA,aAAqB,2BACb;AACR,QAAM,YAAY,KAAK,KAAK,aAAa,KAAK;AAC9C,MAAI,cAAc,SAAS,WAAW;AACpC,WAAO;AAAA,EACT;AAEA,SAAO,oBAAoB,eAAe,UAAU;AACtD;AAKO,MAAM,uBAAuB;AAAA,EACzB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAET,YAAY,QAQT;AACD,SAAK,YAAY,OAAO;AACxB,SAAK,gBAAgB,OAAO,iBAAiB;AAC7C,SAAK,qBAAqB,OAAO,sBAAsB;AACvD,SAAK,iBAAiB,OAAO,kBAAkB;AAC/C,SAAK,cAAc,OAAO;AAC1B,SAAK,gBAAgB,OAAO;AAC5B,SAAK,iBAAiB,OAAO;AAC7B,SAAK,cAAc,KAAK,gBAAgB,oBAAoB,KAAK,aAAa,IAAI;AAAA,EACpF;AAAA,EAEA,OAAO,UAAkC;AACvC,WAAO,IAAI,uBAAuB,EAAE,WAAW,EAAE,CAAC;AAAA,EACpD;AACF;","names":["InterruptionEventType"]}

package/dist/inference/llm.cjs

@@ -132,6 +132,7 @@ class LLMStream extends llm.LLMStream {
   toolIndex;
   fncName;
   fncRawArguments;
+  toolExtra;
   constructor(llm2, {
     model,
     provider,
@@ -157,6 +158,7 @@ class LLMStream extends llm.LLMStream {
     var _a;
     let retryable = true;
     this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = void 0;
+    this.toolExtra = void 0;
     try {
       const messages = await this.chatCtx.toProviderFormat(
         this.providerFmt
@@ -265,12 +267,15 @@ class LLMStream extends llm.LLMStream {
         if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {
           callChunk = this.createRunningToolCallChunk(id, delta);
           this.toolCallId = this.fncName = this.fncRawArguments = void 0;
+          this.toolExtra = void 0;
         }
         if (tool.function.name) {
           this.toolIndex = tool.index;
           this.toolCallId = tool.id;
           this.fncName = tool.function.name;
           this.fncRawArguments = tool.function.arguments || "";
+          this.toolExtra = // eslint-disable-next-line @typescript-eslint/no-explicit-any
+          tool.extra_content ?? void 0;
         } else if (tool.function.arguments) {
           this.fncRawArguments = (this.fncRawArguments || "") + tool.function.arguments;
         }
@@ -282,35 +287,57 @@ class LLMStream extends llm.LLMStream {
     if (choice.finish_reason && ["tool_calls", "stop"].includes(choice.finish_reason) && this.toolCallId !== void 0) {
       const callChunk = this.createRunningToolCallChunk(id, delta);
       this.toolCallId = this.fncName = this.fncRawArguments = void 0;
+      this.toolExtra = void 0;
       return callChunk;
     }
-    if (!delta.content) {
+    const deltaExtra = (
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      delta.extra_content ?? void 0
+    );
+    if (!delta.content && !deltaExtra) {
       return void 0;
     }
     return {
       id,
       delta: {
         role: "assistant",
-        content: delta.content
+        content: delta.content || void 0,
+        extra: deltaExtra
       }
     };
   }
   createRunningToolCallChunk(id, delta) {
+    const toolExtra = this.toolExtra ? { ...this.toolExtra } : {};
+    const thoughtSignature = this.extractThoughtSignature(toolExtra);
+    const deltaExtra = (
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      delta.extra_content ?? void 0
+    );
     return {
       id,
       delta: {
         role: "assistant",
         content: delta.content || void 0,
+        extra: deltaExtra,
         toolCalls: [
           llm.FunctionCall.create({
             callId: this.toolCallId || "",
             name: this.fncName || "",
-            args: this.fncRawArguments || ""
+            args: this.fncRawArguments || "",
+            extra: toolExtra,
+            thoughtSignature
           })
         ]
       }
     };
   }
+  extractThoughtSignature(extra) {
+    const googleExtra = extra == null ? void 0 : extra.google;
+    if (googleExtra && typeof googleExtra === "object") {
+      return googleExtra.thoughtSignature || googleExtra.thought_signature;
+    }
+    return void 0;
+  }
 }
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {

package/dist/inference/llm.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/inference/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport OpenAI from 'openai';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  type Expand,\n  toError,\n} from '../index.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { type AnyString, createAccessToken } from './utils.js';\n\nconst DEFAULT_BASE_URL = 'https://agent-gateway.livekit.cloud/v1';\n\nexport type OpenAIModels =\n  | 'openai/gpt-5'\n  | 'openai/gpt-5-mini'\n  | 'openai/gpt-5-nano'\n  | 'openai/gpt-4.1'\n  | 'openai/gpt-4.1-mini'\n  | 'openai/gpt-4.1-nano'\n  | 'openai/gpt-4o'\n  | 'openai/gpt-4o-mini'\n  | 'openai/gpt-oss-120b';\n\nexport type GoogleModels = 'google/gemini-2.0-flash-lite';\n\nexport type QwenModels = 'qwen/qwen3-235b-a22b-instruct';\n\nexport type KimiModels = 'moonshotai/kimi-k2-instruct';\n\nexport type DeepSeekModels = 'deepseek-ai/deepseek-v3';\n\ntype ChatCompletionPredictionContentParam =\n  Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;\ntype WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;\ntype ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;\ntype Verbosity = 'low' | 'medium' | 'high';\n\nexport interface ChatCompletionOptions extends Record<string, unknown> {\n  frequency_penalty?: number;\n  logit_bias?: Record<string, number>;\n  logprobs?: boolean;\n  max_completion_tokens?: number;\n  max_tokens?: number;\n  metadata?: Record<string, string>;\n  modalities?: Array<'text' | 'audio'>;\n  n?: number;\n  parallel_tool_calls?: boolean;\n  prediction?: ChatCompletionPredictionContentParam | null;\n  presence_penalty?: number;\n  prompt_cache_key?: string;\n  reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n  safety_identifier?: string;\n  seed?: number;\n  service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';\n  stop?: string | string[];\n  store?: boolean;\n  temperature?: number;\n  top_logprobs?: number;\n  top_p?: number;\n  user?: string;\n  verbosity?: Verbosity;\n  web_search_options?: WebSearchOptions;\n\n  // livekit-typed arguments\n  tool_choice?: ToolChoice;\n  // TODO(brian): support response format\n  // response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams['response_format']\n}\n\nexport type LLMModels =\n  | OpenAIModels\n  | GoogleModels\n  | QwenModels\n  | KimiModels\n  | DeepSeekModels\n  | AnyString;\n\nexport interface InferenceLLMOptions {\n  model: LLMModels;\n  provider?: string;\n  baseURL: string;\n  apiKey: string;\n  apiSecret: string;\n  modelOptions: ChatCompletionOptions;\n  strictToolSchema?: boolean;\n}\n\nexport interface GatewayOptions {\n  apiKey: string;\n  apiSecret: string;\n}\n\n/**\n * Livekit Cloud Inference LLM\n */\nexport class LLM extends llm.LLM {\n  private client: OpenAI;\n  private opts: InferenceLLMOptions;\n\n  constructor(opts: {\n    model: LLMModels;\n    provider?: string;\n    baseURL?: string;\n    apiKey?: string;\n    apiSecret?: string;\n    modelOptions?: InferenceLLMOptions['modelOptions'];\n    strictToolSchema?: boolean;\n  }) {\n    super();\n\n    const {\n      model,\n      provider,\n      baseURL,\n      apiKey,\n      apiSecret,\n      modelOptions,\n      strictToolSchema = false,\n    } = opts;\n\n    const lkBaseURL = baseURL || process.env.LIVEKIT_INFERENCE_URL || DEFAULT_BASE_URL;\n    const lkApiKey = apiKey || process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY;\n    if (!lkApiKey) {\n      throw new Error('apiKey is required: pass apiKey or set LIVEKIT_API_KEY');\n    }\n\n    const lkApiSecret =\n      apiSecret || process.env.LIVEKIT_INFERENCE_API_SECRET || process.env.LIVEKIT_API_SECRET;\n    if (!lkApiSecret) {\n      throw new Error('apiSecret is required: pass apiSecret or set LIVEKIT_API_SECRET');\n    }\n\n    this.opts = {\n      model,\n      provider,\n      baseURL: lkBaseURL,\n      apiKey: lkApiKey,\n      apiSecret: lkApiSecret,\n      modelOptions: modelOptions || {},\n      strictToolSchema,\n    };\n\n    this.client = new OpenAI({\n      baseURL: this.opts.baseURL,\n      apiKey: '', // leave a temporary empty string to avoid OpenAI complain about missing key\n    });\n  }\n\n  label(): string {\n    return 'inference.LLM';\n  }\n\n  get model(): string {\n    return this.opts.model;\n  }\n\n  static fromModelString(modelString: string): LLM {\n    return new LLM({ model: modelString });\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    // TODO(AJS-270): Add response_format parameter support\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    // TODO(AJS-270): Add responseFormat parameter\n    extraKwargs?: Record<string, unknown>;\n  }): LLMStream {\n    let modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined\n        ? parallelToolCalls\n        : this.opts.modelOptions.parallel_tool_calls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined\n        ? toolChoice\n        : (this.opts.modelOptions.tool_choice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice as ToolChoice;\n    }\n\n    // TODO(AJS-270): Add response_format support here\n\n    modelOptions = { ...modelOptions, ...this.opts.modelOptions };\n\n    return new LLMStream(this, {\n      model: this.opts.model,\n      provider: this.opts.provider,\n      client: this.client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.opts.strictToolSchema ?? false, // default to false if not set\n      gatewayOptions: {\n        apiKey: this.opts.apiKey,\n        apiSecret: this.opts.apiSecret,\n      },\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  private model: LLMModels;\n  private provider?: string;\n  private providerFmt: llm.ProviderFormat;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n\n  private gatewayOptions?: GatewayOptions;\n  private toolCallId?: string;\n  private toolIndex?: number;\n  private fncName?: string;\n  private fncRawArguments?: string;\n\n  constructor(\n    llm: LLM,\n    {\n      model,\n      provider,\n      client,\n      chatCtx,\n      toolCtx,\n      gatewayOptions,\n      connOptions,\n      modelOptions,\n      providerFmt,\n      strictToolSchema,\n    }: {\n      model: LLMModels;\n      provider?: string;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      gatewayOptions?: GatewayOptions;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      providerFmt?: llm.ProviderFormat;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.client = client;\n    this.gatewayOptions = gatewayOptions;\n    this.provider = provider;\n    this.providerFmt = providerFmt || 'openai';\n    this.modelOptions = modelOptions;\n    this.model = model;\n    this.strictToolSchema = strictToolSchema;\n  }\n\n  protected async run(): Promise<void> {\n    // current function call that we're waiting for full completion (args are streamed)\n    // (defined inside the run method to make sure the state is reset for each run/attempt)\n    let retryable = true;\n    this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = undefined;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        this.providerFmt,\n      )) as OpenAI.ChatCompletionMessageParam[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              function: {\n                name,\n                description: func.description,\n                parameters: llm.toJsonSchema(\n                  func.parameters,\n                  true,\n                  this.strictToolSchema,\n                ) as unknown as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function']['parameters'],\n              } as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function'],\n            };\n\n            if (this.strictToolSchema) {\n              oaiParams.function.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      // Dynamically set the access token for the LiveKit Agent Gateway API\n      if (this.gatewayOptions) {\n        this.client.apiKey = await createAccessToken(\n          this.gatewayOptions.apiKey,\n          this.gatewayOptions.apiSecret,\n        );\n      }\n\n      if (this.provider) {\n        const extraHeaders = requestOptions.extra_headers\n          ? (requestOptions.extra_headers as Record<string, string>)\n          : {};\n        extraHeaders['X-LiveKit-Inference-Provider'] = this.provider;\n        requestOptions.extra_headers = extraHeaders;\n      }\n\n      const stream = await this.client.chat.completions.create(\n        {\n          model: this.model,\n          messages,\n          tools,\n          stream: true,\n          stream_options: { include_usage: true },\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const chunk of stream) {\n        for (const choice of chunk.choices) {\n          if (this.abortController.signal.aborted) {\n            break;\n          }\n          const chatChunk = this.parseChoice(chunk.id, choice);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usage) {\n          const usage = chunk.usage;\n          retryable = false;\n          this.queue.put({\n            id: chunk.id,\n            usage: {\n              completionTokens: usage.completion_tokens,\n              promptTokens: usage.prompt_tokens,\n              promptCachedTokens: usage.prompt_tokens_details?.cached_tokens || 0,\n              totalTokens: usage.total_tokens,\n            },\n          });\n        }\n      }\n    } catch (error) {\n      if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    }\n  }\n\n  private parseChoice(\n    id: string,\n    choice: OpenAI.ChatCompletionChunk.Choice,\n  ): llm.ChatChunk | undefined {\n    const delta = choice.delta;\n\n    // https://github.com/livekit/agents/issues/688\n    // the delta can be None when using Azure OpenAI (content filtering)\n    if (delta === undefined) return undefined;\n\n    if (delta.tool_calls) {\n      // check if we have functions to calls\n      for (const tool of delta.tool_calls) {\n        if (!tool.function) {\n          continue; // oai may add other tools in the future\n        }\n\n        /**\n         * The way OpenAI streams tool calls is a bit tricky.\n         *\n         * For any new tool call, it first emits a delta tool call with id, and function name,\n         * the rest of the delta chunks will only stream the remaining arguments string,\n         * until a new tool call is started or the tool call is finished.\n         * See below for an example.\n         *\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role='assistant', tool_calls=None), finish_reason=None, index=0, logprobs=None)\n         * [ChoiceDeltaToolCall(index=0, id='call_LaVeHWUHpef9K1sd5UO8TtLg', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"P', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='aris\\}', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id='call_ThU4OmMdQXnnVmpXGOCknXIB', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"T', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='okyo', name=None), type=None)]\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role=None, tool_calls=None), finish_reason='tool_calls', index=0, logprobs=None)\n         */\n        let callChunk: llm.ChatChunk | undefined;\n        // If we have a previous tool call and this is a new one, emit the previous\n        if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {\n          callChunk = this.createRunningToolCallChunk(id, delta);\n          this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n        }\n\n        // Start or continue building the current tool call\n        if (tool.function.name) {\n          this.toolIndex = tool.index;\n          this.toolCallId = tool.id;\n          this.fncName = tool.function.name;\n          this.fncRawArguments = tool.function.arguments || '';\n        } else if (tool.function.arguments) {\n          this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;\n        }\n\n        if (callChunk) {\n          return callChunk;\n        }\n      }\n    }\n\n    // If we're done with tool calls, emit the final one\n    if (\n      choice.finish_reason &&\n      ['tool_calls', 'stop'].includes(choice.finish_reason) &&\n      this.toolCallId !== undefined\n    ) {\n      const callChunk = this.createRunningToolCallChunk(id, delta);\n      this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n      return callChunk;\n    }\n\n    // Regular content message\n    if (!delta.content) {\n      return undefined;\n    }\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content,\n      },\n    };\n  }\n\n  private createRunningToolCallChunk(\n    id: string,\n    delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,\n  ): llm.ChatChunk {\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        toolCalls: [\n          llm.FunctionCall.create({\n            callId: this.toolCallId || '',\n            name: this.fncName || '',\n            args: this.fncRawArguments || '',\n          }),\n        ],\n      },\n    };\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,oBAAmB;AACnB,eAOO;AACP,UAAqB;AAErB,mBAAkD;AAElD,MAAM,mBAAmB;AAqFlB,MAAM,YAAY,IAAI,IAAI;AAAA,EACvB;AAAA,EACA;AAAA,EAER,YAAY,MAQT;AACD,UAAM;AAEN,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,mBAAmB;AAAA,IACrB,IAAI;AAEJ,UAAM,YAAY,WAAW,QAAQ,IAAI,yBAAyB;AAClE,UAAM,WAAW,UAAU,QAAQ,IAAI,6BAA6B,QAAQ,IAAI;AAChF,QAAI,CAAC,UAAU;AACb,YAAM,IAAI,MAAM,wDAAwD;AAAA,IAC1E;AAEA,UAAM,cACJ,aAAa,QAAQ,IAAI,gCAAgC,QAAQ,IAAI;AACvE,QAAI,CAAC,aAAa;AAChB,YAAM,IAAI,MAAM,iEAAiE;AAAA,IACnF;AAEA,SAAK,OAAO;AAAA,MACV;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,cAAc,gBAAgB,CAAC;AAAA,MAC/B;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,cAAAA,QAAO;AAAA,MACvB,SAAS,KAAK,KAAK;AAAA,MACnB,QAAQ;AAAA;AAAA,IACV,CAAC;AAAA,EACH;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAEA,OAAO,gBAAgB,aAA0B;AAC/C,WAAO,IAAI,IAAI,EAAE,OAAO,YAAY,CAAC;AAAA,EACvC;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA;AAAA,IAEA;AAAA,EACF,GAQc;AACZ,QAAI,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAErE,wBACE,sBAAsB,SAClB,oBACA,KAAK,KAAK,aAAa;AAE7B,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SACX,aACC,KAAK,KAAK,aAAa;AAE9B,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAIA,mBAAe,EAAE,GAAG,cAAc,GAAG,KAAK,KAAK,aAAa;AAE5D,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,OAAO,KAAK,KAAK;AAAA,MACjB,UAAU,KAAK,KAAK;AAAA,MACpB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,KAAK,oBAAoB;AAAA;AAAA,MAChD,gBAAgB;AAAA,QACd,QAAQ,KAAK,KAAK;AAAA,QAClB,WAAW,KAAK,KAAK;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EACnC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEC,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAYA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,SAAS;AACd,SAAK,iBAAiB;AACtB,SAAK,WAAW;AAChB,SAAK,cAAc,eAAe;AAClC,SAAK,eAAe;AACpB,SAAK,QAAQ;AACb,SAAK,mBAAmB;AAAA,EAC1B;AAAA,EAEA,MAAgB,MAAqB;AAlRvC;AAqRI,QAAI,YAAY;AAChB,SAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB,KAAK,YAAY;AAEzE,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC,KAAK;AAAA,MACP;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN,UAAU;AAAA,YACR;AAAA,YACA,aAAa,KAAK;AAAA,YAClB,YAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL;AAAA,cACA,KAAK;AAAA,YACP;AAAA,UACF;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS,SAAS;AAAA,QAC9B;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAGA,UAAI,KAAK,gBAAgB;AACvB,aAAK,OAAO,SAAS,UAAM;AAAA,UACzB,KAAK,eAAe;AAAA,UACpB,KAAK,eAAe;AAAA,QACtB;AAAA,MACF;AAEA,UAAI,KAAK,UAAU;AACjB,cAAM,eAAe,eAAe,gBAC/B,eAAe,gBAChB,CAAC;AACL,qBAAa,8BAA8B,IAAI,KAAK;AACpD,uBAAe,gBAAgB;AAAA,MACjC;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY;AAAA,QAChD;AAAA,UACE,OAAO,KAAK;AAAA,UACZ;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,gBAAgB,EAAE,eAAe,KAAK;AAAA,UACtC,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,mBAAW,UAAU,MAAM,SAAS;AAClC,cAAI,KAAK,gBAAgB,OAAO,SAAS;AACvC;AAAA,UACF;AACA,gBAAM,YAAY,KAAK,YAAY,MAAM,IAAI,MAAM;AACnD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,OAAO;AACf,gBAAM,QAAQ,MAAM;AACpB,sBAAY;AACZ,eAAK,MAAM,IAAI;AAAA,YACb,IAAI,MAAM;AAAA,YACV,OAAO;AAAA,cACL,kBAAkB,MAAM;AAAA,cACxB,cAAc,MAAM;AAAA,cACpB,sBAAoB,WAAM,0BAAN,mBAA6B,kBAAiB;AAAA,cAClE,aAAa,MAAM;AAAA,YACrB;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,cAAAD,QAAO,2BAA2B;AACrD,cAAM,IAAI,yBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,cAAAA,QAAO,UAAU;AAC3C,cAAM,IAAI,wBAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,4BAAmB;AAAA,UAC3B,aAAS,kBAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,YACN,IACA,QAC2B;AAC3B,UAAM,QAAQ,OAAO;AAIrB,QAAI,UAAU,OAAW,QAAO;AAEhC,QAAI,MAAM,YAAY;AAEpB,iBAAW,QAAQ,MAAM,YAAY;AACnC,YAAI,CAAC,KAAK,UAAU;AAClB;AAAA,QACF;AAmBA,YAAI;AAEJ,YAAI,KAAK,cAAc,KAAK,MAAM,KAAK,UAAU,KAAK,WAAW;AAC/D,sBAAY,KAAK,2BAA2B,IAAI,KAAK;AACrD,eAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AAAA,QAC1D;AAGA,YAAI,KAAK,SAAS,MAAM;AACtB,eAAK,YAAY,KAAK;AACtB,eAAK,aAAa,KAAK;AACvB,eAAK,UAAU,KAAK,SAAS;AAC7B,eAAK,kBAAkB,KAAK,SAAS,aAAa;AAAA,QACpD,WAAW,KAAK,SAAS,WAAW;AAClC,eAAK,mBAAmB,KAAK,mBAAmB,MAAM,KAAK,SAAS;AAAA,QACtE;AAEA,YAAI,WAAW;AACb,iBAAO;AAAA,QACT;AAAA,MACF;AAAA,IACF;AAGA,QACE,OAAO,iBACP,CAAC,cAAc,MAAM,EAAE,SAAS,OAAO,aAAa,KACpD,KAAK,eAAe,QACpB;AACA,YAAM,YAAY,KAAK,2BAA2B,IAAI,KAAK;AAC3D,WAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AACxD,aAAO;AAAA,IACT;AAGA,QAAI,CAAC,MAAM,SAAS;AAClB,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM;AAAA,MACjB;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BACN,IACA,OACe;AACf,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,WAAW;AAAA,UACT,IAAI,aAAa,OAAO;AAAA,YACtB,QAAQ,KAAK,cAAc;AAAA,YAC3B,MAAM,KAAK,WAAW;AAAA,YACtB,MAAM,KAAK,mBAAmB;AAAA,UAChC,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;","names":["OpenAI","llm"]}
+{"version":3,"sources":["../../src/inference/llm.ts"],"sourcesContent":["// SPDX-FileCopyrightText: 2025 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport OpenAI from 'openai';\nimport {\n  APIConnectionError,\n  APIStatusError,\n  APITimeoutError,\n  DEFAULT_API_CONNECT_OPTIONS,\n  type Expand,\n  toError,\n} from '../index.js';\nimport * as llm from '../llm/index.js';\nimport type { APIConnectOptions } from '../types.js';\nimport { type AnyString, createAccessToken } from './utils.js';\n\nconst DEFAULT_BASE_URL = 'https://agent-gateway.livekit.cloud/v1';\n\nexport type OpenAIModels =\n  | 'openai/gpt-5'\n  | 'openai/gpt-5-mini'\n  | 'openai/gpt-5-nano'\n  | 'openai/gpt-4.1'\n  | 'openai/gpt-4.1-mini'\n  | 'openai/gpt-4.1-nano'\n  | 'openai/gpt-4o'\n  | 'openai/gpt-4o-mini'\n  | 'openai/gpt-oss-120b';\n\nexport type GoogleModels =\n  | 'google/gemini-3-pro-preview'\n  | 'google/gemini-3-flash-preview'\n  | 'google/gemini-2.5-pro'\n  | 'google/gemini-2.5-flash'\n  | 'google/gemini-2.5-flash-lite'\n  | 'google/gemini-2.0-flash'\n  | 'google/gemini-2.0-flash-lite';\n\nexport type QwenModels = 'qwen/qwen3-235b-a22b-instruct';\n\nexport type KimiModels = 'moonshotai/kimi-k2-instruct';\n\nexport type DeepSeekModels = 'deepseek-ai/deepseek-v3';\n\ntype ChatCompletionPredictionContentParam =\n  Expand<OpenAI.Chat.Completions.ChatCompletionPredictionContent>;\ntype WebSearchOptions = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams.WebSearchOptions>;\ntype ToolChoice = Expand<OpenAI.Chat.Completions.ChatCompletionCreateParams['tool_choice']>;\ntype Verbosity = 'low' | 'medium' | 'high';\n\nexport interface ChatCompletionOptions extends Record<string, unknown> {\n  frequency_penalty?: number;\n  logit_bias?: Record<string, number>;\n  logprobs?: boolean;\n  max_completion_tokens?: number;\n  max_tokens?: number;\n  metadata?: Record<string, string>;\n  modalities?: Array<'text' | 'audio'>;\n  n?: number;\n  parallel_tool_calls?: boolean;\n  prediction?: ChatCompletionPredictionContentParam | null;\n  presence_penalty?: number;\n  prompt_cache_key?: string;\n  reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';\n  safety_identifier?: string;\n  seed?: number;\n  service_tier?: 'auto' | 'default' | 'flex' | 'scale' | 'priority';\n  stop?: string | string[];\n  store?: boolean;\n  temperature?: number;\n  top_logprobs?: number;\n  top_p?: number;\n  user?: string;\n  verbosity?: Verbosity;\n  web_search_options?: WebSearchOptions;\n\n  // livekit-typed arguments\n  tool_choice?: ToolChoice;\n  // TODO(brian): support response format\n  // response_format?: OpenAI.Chat.Completions.ChatCompletionCreateParams['response_format']\n}\n\nexport type LLMModels =\n  | OpenAIModels\n  | GoogleModels\n  | QwenModels\n  | KimiModels\n  | DeepSeekModels\n  | AnyString;\n\nexport interface InferenceLLMOptions {\n  model: LLMModels;\n  provider?: string;\n  baseURL: string;\n  apiKey: string;\n  apiSecret: string;\n  modelOptions: ChatCompletionOptions;\n  strictToolSchema?: boolean;\n}\n\nexport interface GatewayOptions {\n  apiKey: string;\n  apiSecret: string;\n}\n\n/**\n * Livekit Cloud Inference LLM\n */\nexport class LLM extends llm.LLM {\n  private client: OpenAI;\n  private opts: InferenceLLMOptions;\n\n  constructor(opts: {\n    model: LLMModels;\n    provider?: string;\n    baseURL?: string;\n    apiKey?: string;\n    apiSecret?: string;\n    modelOptions?: InferenceLLMOptions['modelOptions'];\n    strictToolSchema?: boolean;\n  }) {\n    super();\n\n    const {\n      model,\n      provider,\n      baseURL,\n      apiKey,\n      apiSecret,\n      modelOptions,\n      strictToolSchema = false,\n    } = opts;\n\n    const lkBaseURL = baseURL || process.env.LIVEKIT_INFERENCE_URL || DEFAULT_BASE_URL;\n    const lkApiKey = apiKey || process.env.LIVEKIT_INFERENCE_API_KEY || process.env.LIVEKIT_API_KEY;\n    if (!lkApiKey) {\n      throw new Error('apiKey is required: pass apiKey or set LIVEKIT_API_KEY');\n    }\n\n    const lkApiSecret =\n      apiSecret || process.env.LIVEKIT_INFERENCE_API_SECRET || process.env.LIVEKIT_API_SECRET;\n    if (!lkApiSecret) {\n      throw new Error('apiSecret is required: pass apiSecret or set LIVEKIT_API_SECRET');\n    }\n\n    this.opts = {\n      model,\n      provider,\n      baseURL: lkBaseURL,\n      apiKey: lkApiKey,\n      apiSecret: lkApiSecret,\n      modelOptions: modelOptions || {},\n      strictToolSchema,\n    };\n\n    this.client = new OpenAI({\n      baseURL: this.opts.baseURL,\n      apiKey: '', // leave a temporary empty string to avoid OpenAI complain about missing key\n    });\n  }\n\n  label(): string {\n    return 'inference.LLM';\n  }\n\n  get model(): string {\n    return this.opts.model;\n  }\n\n  static fromModelString(modelString: string): LLM {\n    return new LLM({ model: modelString });\n  }\n\n  chat({\n    chatCtx,\n    toolCtx,\n    connOptions = DEFAULT_API_CONNECT_OPTIONS,\n    parallelToolCalls,\n    toolChoice,\n    // TODO(AJS-270): Add response_format parameter support\n    extraKwargs,\n  }: {\n    chatCtx: llm.ChatContext;\n    toolCtx?: llm.ToolContext;\n    connOptions?: APIConnectOptions;\n    parallelToolCalls?: boolean;\n    toolChoice?: llm.ToolChoice;\n    // TODO(AJS-270): Add responseFormat parameter\n    extraKwargs?: Record<string, unknown>;\n  }): LLMStream {\n    let modelOptions: Record<string, unknown> = { ...(extraKwargs || {}) };\n\n    parallelToolCalls =\n      parallelToolCalls !== undefined\n        ? parallelToolCalls\n        : this.opts.modelOptions.parallel_tool_calls;\n\n    if (toolCtx && Object.keys(toolCtx).length > 0 && parallelToolCalls !== undefined) {\n      modelOptions.parallel_tool_calls = parallelToolCalls;\n    }\n\n    toolChoice =\n      toolChoice !== undefined\n        ? toolChoice\n        : (this.opts.modelOptions.tool_choice as llm.ToolChoice | undefined);\n\n    if (toolChoice) {\n      modelOptions.tool_choice = toolChoice as ToolChoice;\n    }\n\n    // TODO(AJS-270): Add response_format support here\n\n    modelOptions = { ...modelOptions, ...this.opts.modelOptions };\n\n    return new LLMStream(this, {\n      model: this.opts.model,\n      provider: this.opts.provider,\n      client: this.client,\n      chatCtx,\n      toolCtx,\n      connOptions,\n      modelOptions,\n      strictToolSchema: this.opts.strictToolSchema ?? false, // default to false if not set\n      gatewayOptions: {\n        apiKey: this.opts.apiKey,\n        apiSecret: this.opts.apiSecret,\n      },\n    });\n  }\n}\n\nexport class LLMStream extends llm.LLMStream {\n  private model: LLMModels;\n  private provider?: string;\n  private providerFmt: llm.ProviderFormat;\n  private client: OpenAI;\n  private modelOptions: Record<string, unknown>;\n  private strictToolSchema: boolean;\n\n  private gatewayOptions?: GatewayOptions;\n  private toolCallId?: string;\n  private toolIndex?: number;\n  private fncName?: string;\n  private fncRawArguments?: string;\n  private toolExtra?: Record<string, unknown>;\n\n  constructor(\n    llm: LLM,\n    {\n      model,\n      provider,\n      client,\n      chatCtx,\n      toolCtx,\n      gatewayOptions,\n      connOptions,\n      modelOptions,\n      providerFmt,\n      strictToolSchema,\n    }: {\n      model: LLMModels;\n      provider?: string;\n      client: OpenAI;\n      chatCtx: llm.ChatContext;\n      toolCtx?: llm.ToolContext;\n      gatewayOptions?: GatewayOptions;\n      connOptions: APIConnectOptions;\n      modelOptions: Record<string, unknown>;\n      providerFmt?: llm.ProviderFormat;\n      strictToolSchema: boolean;\n    },\n  ) {\n    super(llm, { chatCtx, toolCtx, connOptions });\n    this.client = client;\n    this.gatewayOptions = gatewayOptions;\n    this.provider = provider;\n    this.providerFmt = providerFmt || 'openai';\n    this.modelOptions = modelOptions;\n    this.model = model;\n    this.strictToolSchema = strictToolSchema;\n  }\n\n  protected async run(): Promise<void> {\n    // current function call that we're waiting for full completion (args are streamed)\n    // (defined inside the run method to make sure the state is reset for each run/attempt)\n    let retryable = true;\n    this.toolCallId = this.fncName = this.fncRawArguments = this.toolIndex = undefined;\n    this.toolExtra = undefined;\n\n    try {\n      const messages = (await this.chatCtx.toProviderFormat(\n        this.providerFmt,\n      )) as OpenAI.ChatCompletionMessageParam[];\n\n      const tools = this.toolCtx\n        ? Object.entries(this.toolCtx).map(([name, func]) => {\n            const oaiParams = {\n              type: 'function' as const,\n              function: {\n                name,\n                description: func.description,\n                parameters: llm.toJsonSchema(\n                  func.parameters,\n                  true,\n                  this.strictToolSchema,\n                ) as unknown as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function']['parameters'],\n              } as OpenAI.Chat.Completions.ChatCompletionFunctionTool['function'],\n            };\n\n            if (this.strictToolSchema) {\n              oaiParams.function.strict = true;\n            }\n\n            return oaiParams;\n          })\n        : undefined;\n\n      const requestOptions: Record<string, unknown> = { ...this.modelOptions };\n      if (!tools) {\n        delete requestOptions.tool_choice;\n      }\n\n      // Dynamically set the access token for the LiveKit Agent Gateway API\n      if (this.gatewayOptions) {\n        this.client.apiKey = await createAccessToken(\n          this.gatewayOptions.apiKey,\n          this.gatewayOptions.apiSecret,\n        );\n      }\n\n      if (this.provider) {\n        const extraHeaders = requestOptions.extra_headers\n          ? (requestOptions.extra_headers as Record<string, string>)\n          : {};\n        extraHeaders['X-LiveKit-Inference-Provider'] = this.provider;\n        requestOptions.extra_headers = extraHeaders;\n      }\n\n      const stream = await this.client.chat.completions.create(\n        {\n          model: this.model,\n          messages,\n          tools,\n          stream: true,\n          stream_options: { include_usage: true },\n          ...requestOptions,\n        },\n        {\n          timeout: this.connOptions.timeoutMs,\n        },\n      );\n\n      for await (const chunk of stream) {\n        for (const choice of chunk.choices) {\n          if (this.abortController.signal.aborted) {\n            break;\n          }\n          const chatChunk = this.parseChoice(chunk.id, choice);\n          if (chatChunk) {\n            retryable = false;\n            this.queue.put(chatChunk);\n          }\n        }\n\n        if (chunk.usage) {\n          const usage = chunk.usage;\n          retryable = false;\n          this.queue.put({\n            id: chunk.id,\n            usage: {\n              completionTokens: usage.completion_tokens,\n              promptTokens: usage.prompt_tokens,\n              promptCachedTokens: usage.prompt_tokens_details?.cached_tokens || 0,\n              totalTokens: usage.total_tokens,\n            },\n          });\n        }\n      }\n    } catch (error) {\n      if (error instanceof OpenAI.APIConnectionTimeoutError) {\n        throw new APITimeoutError({ options: { retryable } });\n      } else if (error instanceof OpenAI.APIError) {\n        throw new APIStatusError({\n          message: error.message,\n          options: {\n            statusCode: error.status,\n            body: error.error,\n            requestId: error.requestID,\n            retryable,\n          },\n        });\n      } else {\n        throw new APIConnectionError({\n          message: toError(error).message,\n          options: { retryable },\n        });\n      }\n    }\n  }\n\n  private parseChoice(\n    id: string,\n    choice: OpenAI.ChatCompletionChunk.Choice,\n  ): llm.ChatChunk | undefined {\n    const delta = choice.delta;\n\n    // https://github.com/livekit/agents/issues/688\n    // the delta can be None when using Azure OpenAI (content filtering)\n    if (delta === undefined) return undefined;\n\n    if (delta.tool_calls) {\n      // check if we have functions to calls\n      for (const tool of delta.tool_calls) {\n        if (!tool.function) {\n          continue; // oai may add other tools in the future\n        }\n\n        /**\n         * The way OpenAI streams tool calls is a bit tricky.\n         *\n         * For any new tool call, it first emits a delta tool call with id, and function name,\n         * the rest of the delta chunks will only stream the remaining arguments string,\n         * until a new tool call is started or the tool call is finished.\n         * See below for an example.\n         *\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role='assistant', tool_calls=None), finish_reason=None, index=0, logprobs=None)\n         * [ChoiceDeltaToolCall(index=0, id='call_LaVeHWUHpef9K1sd5UO8TtLg', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"P', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=0, id=None, function=ChoiceDeltaToolCallFunction(arguments='aris\\}', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id='call_ThU4OmMdQXnnVmpXGOCknXIB', function=ChoiceDeltaToolCallFunction(arguments='', name='get_weather'), type='function')]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='\\{\"location\": \"T', name=None), type=None)]\n         * [ChoiceDeltaToolCall(index=1, id=None, function=ChoiceDeltaToolCallFunction(arguments='okyo', name=None), type=None)]\n         * Choice(delta=ChoiceDelta(content=None, function_call=None, refusal=None, role=None, tool_calls=None), finish_reason='tool_calls', index=0, logprobs=None)\n         */\n        let callChunk: llm.ChatChunk | undefined;\n        // If we have a previous tool call and this is a new one, emit the previous\n        if (this.toolCallId && tool.id && tool.index !== this.toolIndex) {\n          callChunk = this.createRunningToolCallChunk(id, delta);\n          this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n          this.toolExtra = undefined;\n        }\n\n        // Start or continue building the current tool call\n        if (tool.function.name) {\n          this.toolIndex = tool.index;\n          this.toolCallId = tool.id;\n          this.fncName = tool.function.name;\n          this.fncRawArguments = tool.function.arguments || '';\n          // Extract extra from tool call (e.g., Google thought signatures)\n          this.toolExtra =\n            // eslint-disable-next-line @typescript-eslint/no-explicit-any\n            ((tool as any).extra_content as Record<string, unknown> | undefined) ?? undefined;\n        } else if (tool.function.arguments) {\n          this.fncRawArguments = (this.fncRawArguments || '') + tool.function.arguments;\n        }\n\n        if (callChunk) {\n          return callChunk;\n        }\n      }\n    }\n\n    // If we're done with tool calls, emit the final one\n    if (\n      choice.finish_reason &&\n      ['tool_calls', 'stop'].includes(choice.finish_reason) &&\n      this.toolCallId !== undefined\n    ) {\n      const callChunk = this.createRunningToolCallChunk(id, delta);\n      this.toolCallId = this.fncName = this.fncRawArguments = undefined;\n      this.toolExtra = undefined;\n      return callChunk;\n    }\n\n    // Extract extra from delta (e.g., Google thought signatures on text parts)\n    const deltaExtra =\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      ((delta as any).extra_content as Record<string, unknown> | undefined) ?? undefined;\n\n    // Regular content message\n    if (!delta.content && !deltaExtra) {\n      return undefined;\n    }\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        extra: deltaExtra,\n      },\n    };\n  }\n\n  private createRunningToolCallChunk(\n    id: string,\n    delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta,\n  ): llm.ChatChunk {\n    const toolExtra = this.toolExtra ? { ...this.toolExtra } : {};\n    const thoughtSignature = this.extractThoughtSignature(toolExtra);\n    const deltaExtra =\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      ((delta as any).extra_content as Record<string, unknown> | undefined) ?? undefined;\n\n    return {\n      id,\n      delta: {\n        role: 'assistant',\n        content: delta.content || undefined,\n        extra: deltaExtra,\n        toolCalls: [\n          llm.FunctionCall.create({\n            callId: this.toolCallId || '',\n            name: this.fncName || '',\n            args: this.fncRawArguments || '',\n            extra: toolExtra,\n            thoughtSignature,\n          }),\n        ],\n      },\n    };\n  }\n\n  private extractThoughtSignature(extra?: Record<string, unknown>): string | undefined {\n    const googleExtra = extra?.google;\n    if (googleExtra && typeof googleExtra === 'object') {\n      // eslint-disable-next-line @typescript-eslint/no-explicit-any\n      return (googleExtra as any).thoughtSignature || (googleExtra as any).thought_signature;\n    }\n    return undefined;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAGA,oBAAmB;AACnB,eAOO;AACP,UAAqB;AAErB,mBAAkD;AAElD,MAAM,mBAAmB;AA4FlB,MAAM,YAAY,IAAI,IAAI;AAAA,EACvB;AAAA,EACA;AAAA,EAER,YAAY,MAQT;AACD,UAAM;AAEN,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,mBAAmB;AAAA,IACrB,IAAI;AAEJ,UAAM,YAAY,WAAW,QAAQ,IAAI,yBAAyB;AAClE,UAAM,WAAW,UAAU,QAAQ,IAAI,6BAA6B,QAAQ,IAAI;AAChF,QAAI,CAAC,UAAU;AACb,YAAM,IAAI,MAAM,wDAAwD;AAAA,IAC1E;AAEA,UAAM,cACJ,aAAa,QAAQ,IAAI,gCAAgC,QAAQ,IAAI;AACvE,QAAI,CAAC,aAAa;AAChB,YAAM,IAAI,MAAM,iEAAiE;AAAA,IACnF;AAEA,SAAK,OAAO;AAAA,MACV;AAAA,MACA;AAAA,MACA,SAAS;AAAA,MACT,QAAQ;AAAA,MACR,WAAW;AAAA,MACX,cAAc,gBAAgB,CAAC;AAAA,MAC/B;AAAA,IACF;AAEA,SAAK,SAAS,IAAI,cAAAA,QAAO;AAAA,MACvB,SAAS,KAAK,KAAK;AAAA,MACnB,QAAQ;AAAA;AAAA,IACV,CAAC;AAAA,EACH;AAAA,EAEA,QAAgB;AACd,WAAO;AAAA,EACT;AAAA,EAEA,IAAI,QAAgB;AAClB,WAAO,KAAK,KAAK;AAAA,EACnB;AAAA,EAEA,OAAO,gBAAgB,aAA0B;AAC/C,WAAO,IAAI,IAAI,EAAE,OAAO,YAAY,CAAC;AAAA,EACvC;AAAA,EAEA,KAAK;AAAA,IACH;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd;AAAA,IACA;AAAA;AAAA,IAEA;AAAA,EACF,GAQc;AACZ,QAAI,eAAwC,EAAE,GAAI,eAAe,CAAC,EAAG;AAErE,wBACE,sBAAsB,SAClB,oBACA,KAAK,KAAK,aAAa;AAE7B,QAAI,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS,KAAK,sBAAsB,QAAW;AACjF,mBAAa,sBAAsB;AAAA,IACrC;AAEA,iBACE,eAAe,SACX,aACC,KAAK,KAAK,aAAa;AAE9B,QAAI,YAAY;AACd,mBAAa,cAAc;AAAA,IAC7B;AAIA,mBAAe,EAAE,GAAG,cAAc,GAAG,KAAK,KAAK,aAAa;AAE5D,WAAO,IAAI,UAAU,MAAM;AAAA,MACzB,OAAO,KAAK,KAAK;AAAA,MACjB,UAAU,KAAK,KAAK;AAAA,MACpB,QAAQ,KAAK;AAAA,MACb;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,kBAAkB,KAAK,KAAK,oBAAoB;AAAA;AAAA,MAChD,gBAAgB;AAAA,QACd,QAAQ,KAAK,KAAK;AAAA,QAClB,WAAW,KAAK,KAAK;AAAA,MACvB;AAAA,IACF,CAAC;AAAA,EACH;AACF;AAEO,MAAM,kBAAkB,IAAI,UAAU;AAAA,EACnC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAER,YACEC,MACA;AAAA,IACE;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,GAYA;AACA,UAAMA,MAAK,EAAE,SAAS,SAAS,YAAY,CAAC;AAC5C,SAAK,SAAS;AACd,SAAK,iBAAiB;AACtB,SAAK,WAAW;AAChB,SAAK,cAAc,eAAe;AAClC,SAAK,eAAe;AACpB,SAAK,QAAQ;AACb,SAAK,mBAAmB;AAAA,EAC1B;AAAA,EAEA,MAAgB,MAAqB;AA1RvC;AA6RI,QAAI,YAAY;AAChB,SAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB,KAAK,YAAY;AACzE,SAAK,YAAY;AAEjB,QAAI;AACF,YAAM,WAAY,MAAM,KAAK,QAAQ;AAAA,QACnC,KAAK;AAAA,MACP;AAEA,YAAM,QAAQ,KAAK,UACf,OAAO,QAAQ,KAAK,OAAO,EAAE,IAAI,CAAC,CAAC,MAAM,IAAI,MAAM;AACjD,cAAM,YAAY;AAAA,UAChB,MAAM;AAAA,UACN,UAAU;AAAA,YACR;AAAA,YACA,aAAa,KAAK;AAAA,YAClB,YAAY,IAAI;AAAA,cACd,KAAK;AAAA,cACL;AAAA,cACA,KAAK;AAAA,YACP;AAAA,UACF;AAAA,QACF;AAEA,YAAI,KAAK,kBAAkB;AACzB,oBAAU,SAAS,SAAS;AAAA,QAC9B;AAEA,eAAO;AAAA,MACT,CAAC,IACD;AAEJ,YAAM,iBAA0C,EAAE,GAAG,KAAK,aAAa;AACvE,UAAI,CAAC,OAAO;AACV,eAAO,eAAe;AAAA,MACxB;AAGA,UAAI,KAAK,gBAAgB;AACvB,aAAK,OAAO,SAAS,UAAM;AAAA,UACzB,KAAK,eAAe;AAAA,UACpB,KAAK,eAAe;AAAA,QACtB;AAAA,MACF;AAEA,UAAI,KAAK,UAAU;AACjB,cAAM,eAAe,eAAe,gBAC/B,eAAe,gBAChB,CAAC;AACL,qBAAa,8BAA8B,IAAI,KAAK;AACpD,uBAAe,gBAAgB;AAAA,MACjC;AAEA,YAAM,SAAS,MAAM,KAAK,OAAO,KAAK,YAAY;AAAA,QAChD;AAAA,UACE,OAAO,KAAK;AAAA,UACZ;AAAA,UACA;AAAA,UACA,QAAQ;AAAA,UACR,gBAAgB,EAAE,eAAe,KAAK;AAAA,UACtC,GAAG;AAAA,QACL;AAAA,QACA;AAAA,UACE,SAAS,KAAK,YAAY;AAAA,QAC5B;AAAA,MACF;AAEA,uBAAiB,SAAS,QAAQ;AAChC,mBAAW,UAAU,MAAM,SAAS;AAClC,cAAI,KAAK,gBAAgB,OAAO,SAAS;AACvC;AAAA,UACF;AACA,gBAAM,YAAY,KAAK,YAAY,MAAM,IAAI,MAAM;AACnD,cAAI,WAAW;AACb,wBAAY;AACZ,iBAAK,MAAM,IAAI,SAAS;AAAA,UAC1B;AAAA,QACF;AAEA,YAAI,MAAM,OAAO;AACf,gBAAM,QAAQ,MAAM;AACpB,sBAAY;AACZ,eAAK,MAAM,IAAI;AAAA,YACb,IAAI,MAAM;AAAA,YACV,OAAO;AAAA,cACL,kBAAkB,MAAM;AAAA,cACxB,cAAc,MAAM;AAAA,cACpB,sBAAoB,WAAM,0BAAN,mBAA6B,kBAAiB;AAAA,cAClE,aAAa,MAAM;AAAA,YACrB;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF,SAAS,OAAO;AACd,UAAI,iBAAiB,cAAAD,QAAO,2BAA2B;AACrD,cAAM,IAAI,yBAAgB,EAAE,SAAS,EAAE,UAAU,EAAE,CAAC;AAAA,MACtD,WAAW,iBAAiB,cAAAA,QAAO,UAAU;AAC3C,cAAM,IAAI,wBAAe;AAAA,UACvB,SAAS,MAAM;AAAA,UACf,SAAS;AAAA,YACP,YAAY,MAAM;AAAA,YAClB,MAAM,MAAM;AAAA,YACZ,WAAW,MAAM;AAAA,YACjB;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,OAAO;AACL,cAAM,IAAI,4BAAmB;AAAA,UAC3B,aAAS,kBAAQ,KAAK,EAAE;AAAA,UACxB,SAAS,EAAE,UAAU;AAAA,QACvB,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,YACN,IACA,QAC2B;AAC3B,UAAM,QAAQ,OAAO;AAIrB,QAAI,UAAU,OAAW,QAAO;AAEhC,QAAI,MAAM,YAAY;AAEpB,iBAAW,QAAQ,MAAM,YAAY;AACnC,YAAI,CAAC,KAAK,UAAU;AAClB;AAAA,QACF;AAmBA,YAAI;AAEJ,YAAI,KAAK,cAAc,KAAK,MAAM,KAAK,UAAU,KAAK,WAAW;AAC/D,sBAAY,KAAK,2BAA2B,IAAI,KAAK;AACrD,eAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AACxD,eAAK,YAAY;AAAA,QACnB;AAGA,YAAI,KAAK,SAAS,MAAM;AACtB,eAAK,YAAY,KAAK;AACtB,eAAK,aAAa,KAAK;AACvB,eAAK,UAAU,KAAK,SAAS;AAC7B,eAAK,kBAAkB,KAAK,SAAS,aAAa;AAElD,eAAK;AAAA,UAED,KAAa,iBAAyD;AAAA,QAC5E,WAAW,KAAK,SAAS,WAAW;AAClC,eAAK,mBAAmB,KAAK,mBAAmB,MAAM,KAAK,SAAS;AAAA,QACtE;AAEA,YAAI,WAAW;AACb,iBAAO;AAAA,QACT;AAAA,MACF;AAAA,IACF;AAGA,QACE,OAAO,iBACP,CAAC,cAAc,MAAM,EAAE,SAAS,OAAO,aAAa,KACpD,KAAK,eAAe,QACpB;AACA,YAAM,YAAY,KAAK,2BAA2B,IAAI,KAAK;AAC3D,WAAK,aAAa,KAAK,UAAU,KAAK,kBAAkB;AACxD,WAAK,YAAY;AACjB,aAAO;AAAA,IACT;AAGA,UAAM;AAAA;AAAA,MAEF,MAAc,iBAAyD;AAAA;AAG3E,QAAI,CAAC,MAAM,WAAW,CAAC,YAAY;AACjC,aAAO;AAAA,IACT;AAEA,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,OAAO;AAAA,MACT;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BACN,IACA,OACe;AACf,UAAM,YAAY,KAAK,YAAY,EAAE,GAAG,KAAK,UAAU,IAAI,CAAC;AAC5D,UAAM,mBAAmB,KAAK,wBAAwB,SAAS;AAC/D,UAAM;AAAA;AAAA,MAEF,MAAc,iBAAyD;AAAA;AAE3E,WAAO;AAAA,MACL;AAAA,MACA,OAAO;AAAA,QACL,MAAM;AAAA,QACN,SAAS,MAAM,WAAW;AAAA,QAC1B,OAAO;AAAA,QACP,WAAW;AAAA,UACT,IAAI,aAAa,OAAO;AAAA,YACtB,QAAQ,KAAK,cAAc;AAAA,YAC3B,MAAM,KAAK,WAAW;AAAA,YACtB,MAAM,KAAK,mBAAmB;AAAA,YAC9B,OAAO;AAAA,YACP;AAAA,UACF,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,wBAAwB,OAAqD;AACnF,UAAM,cAAc,+BAAO;AAC3B,QAAI,eAAe,OAAO,gBAAgB,UAAU;AAElD,aAAQ,YAAoB,oBAAqB,YAAoB;AAAA,IACvE;AACA,WAAO;AAAA,EACT;AACF;","names":["OpenAI","llm"]}

package/dist/inference/llm.d.cts

@@ -4,7 +4,7 @@ import * as llm from '../llm/index.js';
 import type { APIConnectOptions } from '../types.js';
 import { type AnyString } from './utils.js';
 export type OpenAIModels = 'openai/gpt-5' | 'openai/gpt-5-mini' | 'openai/gpt-5-nano' | 'openai/gpt-4.1' | 'openai/gpt-4.1-mini' | 'openai/gpt-4.1-nano' | 'openai/gpt-4o' | 'openai/gpt-4o-mini' | 'openai/gpt-oss-120b';
-export type GoogleModels = 'google/gemini-2.0-flash-lite';
+export type GoogleModels = 'google/gemini-3-pro-preview' | 'google/gemini-3-flash-preview' | 'google/gemini-2.5-pro' | 'google/gemini-2.5-flash' | 'google/gemini-2.5-flash-lite' | 'google/gemini-2.0-flash' | 'google/gemini-2.0-flash-lite';
 export type QwenModels = 'qwen/qwen3-235b-a22b-instruct';
 export type KimiModels = 'moonshotai/kimi-k2-instruct';
 export type DeepSeekModels = 'deepseek-ai/deepseek-v3';
@@ -92,6 +92,7 @@ export declare class LLMStream extends llm.LLMStream {
     private toolIndex?;
     private fncName?;
     private fncRawArguments?;
+    private toolExtra?;
     constructor(llm: LLM, { model, provider, client, chatCtx, toolCtx, gatewayOptions, connOptions, modelOptions, providerFmt, strictToolSchema, }: {
         model: LLMModels;
         provider?: string;
@@ -107,6 +108,7 @@ export declare class LLMStream extends llm.LLMStream {
     protected run(): Promise<void>;
     private parseChoice;
     private createRunningToolCallChunk;
+    private extractThoughtSignature;
 }
 export {};
 //# sourceMappingURL=llm.d.ts.map

package/dist/inference/llm.d.ts

@@ -4,7 +4,7 @@ import * as llm from '../llm/index.js';
 import type { APIConnectOptions } from '../types.js';
 import { type AnyString } from './utils.js';
 export type OpenAIModels = 'openai/gpt-5' | 'openai/gpt-5-mini' | 'openai/gpt-5-nano' | 'openai/gpt-4.1' | 'openai/gpt-4.1-mini' | 'openai/gpt-4.1-nano' | 'openai/gpt-4o' | 'openai/gpt-4o-mini' | 'openai/gpt-oss-120b';
-export type GoogleModels = 'google/gemini-2.0-flash-lite';
+export type GoogleModels = 'google/gemini-3-pro-preview' | 'google/gemini-3-flash-preview' | 'google/gemini-2.5-pro' | 'google/gemini-2.5-flash' | 'google/gemini-2.5-flash-lite' | 'google/gemini-2.0-flash' | 'google/gemini-2.0-flash-lite';
 export type QwenModels = 'qwen/qwen3-235b-a22b-instruct';
 export type KimiModels = 'moonshotai/kimi-k2-instruct';
 export type DeepSeekModels = 'deepseek-ai/deepseek-v3';
@@ -92,6 +92,7 @@ export declare class LLMStream extends llm.LLMStream {
     private toolIndex?;
     private fncName?;
     private fncRawArguments?;
+    private toolExtra?;
     constructor(llm: LLM, { model, provider, client, chatCtx, toolCtx, gatewayOptions, connOptions, modelOptions, providerFmt, strictToolSchema, }: {
         model: LLMModels;
         provider?: string;
@@ -107,6 +108,7 @@ export declare class LLMStream extends llm.LLMStream {
     protected run(): Promise<void>;
     private parseChoice;
     private createRunningToolCallChunk;
+    private extractThoughtSignature;
 }
 export {};
 //# sourceMappingURL=llm.d.ts.map