@juspay/neurolink 9.71.0 → 9.73.0

package/dist/voice/livekit/roomContext.js

@@ -0,0 +1,56 @@
+/**
+ * Per-call context from LiveKit room metadata.
+ *
+ * The manager (e.g. a Lighthouse `/start` endpoint) pre-creates the room with
+ * `base64(JSON({ authToken, mcpContext }))` metadata, built from the caller's
+ * session. The worker reads it on join — nothing per-call comes from worker env.
+ * Returns the MCP `x-auth-token` and the base64(JSON) `x-context` the server
+ * expects.
+ *
+ * The metadata is untrusted input, so it is decoded with a zod schema rather
+ * than a trusted `JSON.parse` cast.
+ *
+ * See docs/features/livekit-voice-agent.md.
+ */
+import { Buffer } from "node:buffer";
+import { z } from "zod";
+import { logger } from "../../utils/logger.js";
+/** Shape the manager writes into room metadata. `mcpContext` is opaque here. */
+const roomMetadataSchema = z.object({
+    authToken: z.string().optional(),
+    mcpContext: z.unknown().optional(),
+});
+/** Decode the base64(JSON) metadata string into an `unknown`, or `undefined`. */
+function decodeBase64Json(encoded) {
+    try {
+        return JSON.parse(Buffer.from(encoded, "base64").toString("utf-8"));
+    }
+    catch (error) {
+        logger.error(`[RealtimeVoiceAgent] room metadata is not valid base64 JSON: ${String(error)}`);
+        return undefined;
+    }
+}
+/**
+ * Decode `{ authToken, mcpContext }` from a room's base64(JSON) metadata.
+ *
+ * `authToken` may be empty (demo/guest, where the MCP server gates on the
+ * context's `demoMode`); `xContext` is the re-encoded base64(JSON) of
+ * `mcpContext`, or `""` when no context was supplied or the metadata is invalid.
+ */
+export function readCallContextFromRoom(roomMetadata) {
+    const empty = { authToken: "", xContext: "" };
+    if (!roomMetadata) {
+        logger.warn("[RealtimeVoiceAgent] room has no metadata — MCP auth/context unavailable.");
+        return empty;
+    }
+    const decoded = roomMetadataSchema.safeParse(decodeBase64Json(roomMetadata));
+    if (!decoded.success) {
+        logger.error(`[RealtimeVoiceAgent] room metadata has unexpected shape: ${decoded.error.message}`);
+        return empty;
+    }
+    const { authToken, mcpContext } = decoded.data;
+    const xContext = mcpContext === undefined || mcpContext === null
+        ? ""
+        : Buffer.from(JSON.stringify(mcpContext), "utf-8").toString("base64");
+    return { authToken: authToken ?? "", xContext };
+}

package/dist/voice/livekit/roomDispatch.d.ts

@@ -0,0 +1,24 @@
+/**
+ * LiveKit server-side room operations: create a room with metadata, and
+ * dispatch a named agent to a room.
+ *
+ * Wraps `livekit-server-sdk` (an optional dependency, imported dynamically) so
+ * consumers route all LiveKit *server* calls through `@juspay/neurolink/livekit`
+ * — they never depend on the SDK directly. Mirrors `mintJoinToken`.
+ */
+import type { LiveKitServerCredentials } from "../../types/index.js";
+export declare function createVoiceRoom(req: LiveKitServerCredentials & {
+    room: string;
+    metadata?: string;
+    emptyTimeoutSeconds?: number;
+    departureTimeoutSeconds?: number;
+}): Promise<void>;
+/**
+ * Explicitly dispatch a named agent to a room. The long-lived worker registered
+ * under `agentName` receives the job and forks a child to run the call.
+ */
+export declare function dispatchVoiceAgent(req: LiveKitServerCredentials & {
+    room: string;
+    agentName: string;
+    metadata?: string;
+}): Promise<void>;

package/dist/voice/livekit/roomDispatch.js

@@ -0,0 +1,30 @@
+/**
+ * LiveKit server-side room operations: create a room with metadata, and
+ * dispatch a named agent to a room.
+ *
+ * Wraps `livekit-server-sdk` (an optional dependency, imported dynamically) so
+ * consumers route all LiveKit *server* calls through `@juspay/neurolink/livekit`
+ * — they never depend on the SDK directly. Mirrors `mintJoinToken`.
+ */
+const toHttpUrl = (url) => url.replace(/^ws/, "http");
+export async function createVoiceRoom(req) {
+    const { RoomServiceClient } = await import("livekit-server-sdk");
+    const client = new RoomServiceClient(toHttpUrl(req.url), req.apiKey, req.apiSecret);
+    await client.createRoom({
+        name: req.room,
+        metadata: req.metadata ?? "",
+        emptyTimeout: req.emptyTimeoutSeconds ?? 300,
+        departureTimeout: req.departureTimeoutSeconds ?? 20,
+    });
+}
+/**
+ * Explicitly dispatch a named agent to a room. The long-lived worker registered
+ * under `agentName` receives the job and forks a child to run the call.
+ */
+export async function dispatchVoiceAgent(req) {
+    const { AgentDispatchClient } = await import("livekit-server-sdk");
+    const client = new AgentDispatchClient(toHttpUrl(req.url), req.apiKey, req.apiSecret);
+    await client.createDispatch(req.room, req.agentName, {
+        metadata: req.metadata ?? "",
+    });
+}

package/dist/voice/livekit/schemaSanitizer.d.ts

@@ -0,0 +1,26 @@
+/**
+ * Gemini function-calling schema sanitizer.
+ *
+ * Normalises an MCP tool's JSON Schema into the subset Gemini's function-calling
+ * accepts. `@google/genai`'s processJsonSchema crashes on untyped nodes,
+ * `$ref`/`$defs`, and some `anyOf`/`oneOf` shapes, so we rebuild a clean tree:
+ * every node gets a concrete `type`, unions collapse to their first concrete
+ * branch, and unsupported keywords are dropped.
+ *
+ * Pure, dependency-free, and assertion-free — values arrive as `unknown` and are
+ * narrowed with the `isRecord` guard. Safe to reuse for any Gemini tool path.
+ */
+/**
+ * Rebuild a JSON Schema node into the Gemini-safe subset. Returns a fresh object
+ * with a concrete `type` on every node.
+ */
+export declare function sanitizeSchema(node: unknown): Record<string, unknown>;
+/** Tool parameters must be an object schema; force it and sanitize the tree. */
+export declare function sanitizeToolParameters(schema: unknown): Record<string, unknown>;
+/**
+ * Walk a (sanitized) schema and return the first node the google plugin would
+ * turn into `undefined` — which genai then crashes on. Returns a human-readable
+ * path/reason, or `null` if the schema is safe. After `sanitizeSchema` this
+ * should always be `null`; if not, it names the exact offending path.
+ */
+export declare function findSchemaIssue(node: unknown, pathPrefix?: string): string | null;

package/dist/voice/livekit/schemaSanitizer.js

@@ -0,0 +1,143 @@
+/**
+ * Gemini function-calling schema sanitizer.
+ *
+ * Normalises an MCP tool's JSON Schema into the subset Gemini's function-calling
+ * accepts. `@google/genai`'s processJsonSchema crashes on untyped nodes,
+ * `$ref`/`$defs`, and some `anyOf`/`oneOf` shapes, so we rebuild a clean tree:
+ * every node gets a concrete `type`, unions collapse to their first concrete
+ * branch, and unsupported keywords are dropped.
+ *
+ * Pure, dependency-free, and assertion-free — values arrive as `unknown` and are
+ * narrowed with the `isRecord` guard. Safe to reuse for any Gemini tool path.
+ */
+const GEMINI_TYPES = new Set([
+    "string",
+    "number",
+    "integer",
+    "boolean",
+    "array",
+    "object",
+]);
+function isRecord(value) {
+    return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+/**
+ * The first concrete (non-`"null"`-typed) branch of an `anyOf`/`oneOf`/`allOf`
+ * union, or `undefined` when there is no union to collapse.
+ */
+function firstConcreteUnionBranch(schema) {
+    const union = schema.anyOf ?? schema.oneOf ?? schema.allOf;
+    if (!Array.isArray(union)) {
+        return undefined;
+    }
+    return union.find((branch) => isRecord(branch) && branch.type !== "null");
+}
+function resolveSchemaType(schema) {
+    if (typeof schema.type === "string") {
+        return GEMINI_TYPES.has(schema.type) ? schema.type : "string";
+    }
+    if (Array.isArray(schema.type)) {
+        const named = schema.type.find((entry) => typeof entry === "string" && entry !== "null");
+        if (named !== undefined && GEMINI_TYPES.has(named)) {
+            return named;
+        }
+    }
+    if (isRecord(schema.properties)) {
+        return "object";
+    }
+    if (schema.items !== undefined) {
+        return "array";
+    }
+    return "string";
+}
+function sanitizeObjectMembers(schema, out) {
+    const properties = isRecord(schema.properties) ? schema.properties : {};
+    const sanitizedProperties = {};
+    for (const [key, value] of Object.entries(properties)) {
+        sanitizedProperties[key] = sanitizeSchema(value);
+    }
+    out.properties = sanitizedProperties;
+    if (Array.isArray(schema.required)) {
+        out.required = schema.required.filter((name) => typeof name === "string" && name in sanitizedProperties);
+    }
+    if (Object.keys(sanitizedProperties).length === 0) {
+        out.additionalProperties = true;
+    }
+}
+/**
+ * Rebuild a JSON Schema node into the Gemini-safe subset. Returns a fresh object
+ * with a concrete `type` on every node.
+ */
+export function sanitizeSchema(node) {
+    if (!isRecord(node)) {
+        return { type: "string" };
+    }
+    const out = {};
+    if (typeof node.description === "string") {
+        out.description = node.description;
+    }
+    if (typeof node.type !== "string") {
+        const branch = firstConcreteUnionBranch(node);
+        if (branch !== undefined) {
+            const merged = sanitizeSchema(branch);
+            return out.description
+                ? { ...merged, description: out.description }
+                : merged;
+        }
+    }
+    const type = resolveSchemaType(node);
+    out.type = type;
+    if (Array.isArray(node.enum)) {
+        out.enum = node.enum;
+    }
+    if (type === "object") {
+        sanitizeObjectMembers(node, out);
+    }
+    if (type === "array") {
+        out.items = sanitizeSchema(node.items);
+    }
+    return out;
+}
+/** Tool parameters must be an object schema; force it and sanitize the tree. */
+export function sanitizeToolParameters(schema) {
+    const sanitized = sanitizeSchema(schema ?? {});
+    if (sanitized.type !== "object") {
+        return { type: "object", properties: {}, additionalProperties: true };
+    }
+    return sanitized;
+}
+/**
+ * Walk a (sanitized) schema and return the first node the google plugin would
+ * turn into `undefined` — which genai then crashes on. Returns a human-readable
+ * path/reason, or `null` if the schema is safe. After `sanitizeSchema` this
+ * should always be `null`; if not, it names the exact offending path.
+ */
+export function findSchemaIssue(node, pathPrefix = "$") {
+    if (!isRecord(node)) {
+        return `${pathPrefix}: not an object schema`;
+    }
+    if (typeof node.type !== "string") {
+        return `${pathPrefix}: missing string "type"`;
+    }
+    if (node.type === "object") {
+        const properties = isRecord(node.properties) ? node.properties : undefined;
+        const isEmpty = properties === undefined || Object.keys(properties).length === 0;
+        if (isEmpty &&
+            (node.additionalProperties === undefined ||
+                node.additionalProperties === null)) {
+            return `${pathPrefix}: empty object schema without additionalProperties (plugin → undefined)`;
+        }
+        if (properties !== undefined) {
+            for (const [key, value] of Object.entries(properties)) {
+                const childIssue = findSchemaIssue(value, `${pathPrefix}.${key}`);
+                if (childIssue) {
+                    return childIssue;
+                }
+            }
+        }
+    }
+    if (node.type === "array") {
+        return findSchemaIssue(node.items, `${pathPrefix}[]`);
+    }
+    return null;
+}

package/dist/voice/livekit/vertexAuth.d.ts

@@ -0,0 +1,30 @@
+/**
+ * Vertex authentication helpers for the realtime voice agent.
+ *
+ * The Gemini Live WebSocket authenticates to Vertex via Application Default
+ * Credentials (ADC). These helpers materialise ADC from the split
+ * `GOOGLE_AUTH_*` env fields when no credentials file is configured, and remove
+ * any Gemini Developer API key from the environment so `@google/genai` uses
+ * Vertex/ADC auth (not an API key) for the realtime WebSocket.
+ *
+ * See docs/features/livekit-voice-agent.md.
+ */
+/**
+ * Materialise Vertex ADC from the split `GOOGLE_AUTH_*` env fields.
+ *
+ * The google realtime plugin authenticates Vertex via ADC (it does not accept
+ * inline credentials), so this writes a temp service-account JSON and points
+ * `GOOGLE_APPLICATION_CREDENTIALS` at it — unless ADC is already configured.
+ * No-op when `GOOGLE_APPLICATION_CREDENTIALS` is set or the `GOOGLE_AUTH_*`
+ * fields are absent (auth then relies on ambient ADC).
+ */
+export declare function ensureVertexAdc(): void;
+/**
+ * Force pure Vertex/ADC auth for the Gemini Live WebSocket.
+ *
+ * `@google/genai` 1.52+ uses a Gemini Developer API key for the realtime
+ * WebSocket auth even when `vertexai: true` and project/location are set, which
+ * Vertex rejects at the handshake (WS close 1006). The realtime worker only
+ * ever talks to Vertex, so remove these keys (only affects this process).
+ */
+export declare function clearGeminiApiKeyEnv(): void;

package/dist/voice/livekit/vertexAuth.js

@@ -0,0 +1,72 @@
+/**
+ * Vertex authentication helpers for the realtime voice agent.
+ *
+ * The Gemini Live WebSocket authenticates to Vertex via Application Default
+ * Credentials (ADC). These helpers materialise ADC from the split
+ * `GOOGLE_AUTH_*` env fields when no credentials file is configured, and remove
+ * any Gemini Developer API key from the environment so `@google/genai` uses
+ * Vertex/ADC auth (not an API key) for the realtime WebSocket.
+ *
+ * See docs/features/livekit-voice-agent.md.
+ */
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import { logger } from "../../utils/logger.js";
+/**
+ * Materialise Vertex ADC from the split `GOOGLE_AUTH_*` env fields.
+ *
+ * The google realtime plugin authenticates Vertex via ADC (it does not accept
+ * inline credentials), so this writes a temp service-account JSON and points
+ * `GOOGLE_APPLICATION_CREDENTIALS` at it — unless ADC is already configured.
+ * No-op when `GOOGLE_APPLICATION_CREDENTIALS` is set or the `GOOGLE_AUTH_*`
+ * fields are absent (auth then relies on ambient ADC).
+ */
+export function ensureVertexAdc() {
+    if (process.env.GOOGLE_APPLICATION_CREDENTIALS) {
+        return;
+    }
+    const clientEmail = process.env.GOOGLE_AUTH_CLIENT_EMAIL;
+    const rawPrivateKey = process.env.GOOGLE_AUTH_PRIVATE_KEY;
+    if (!clientEmail || !rawPrivateKey) {
+        logger.warn("[RealtimeVoiceAgent] No GOOGLE_APPLICATION_CREDENTIALS and no GOOGLE_AUTH_* fields — Vertex auth will rely on ambient ADC.");
+        return;
+    }
+    const credentials = {
+        type: process.env.GOOGLE_AUTH_TYPE ?? "service_account",
+        project_id: process.env.GOOGLE_AUTH_BREEZE_PROJECT_ID ??
+            process.env.GOOGLE_CLOUD_PROJECT_ID,
+        private_key_id: process.env.GOOGLE_AUTH_PRIVATE_KEY_ID,
+        private_key: rawPrivateKey.replace(/\\n/g, "\n"),
+        client_email: clientEmail,
+        token_uri: process.env.GOOGLE_AUTH_TOKEN_URI ??
+            "https://oauth2.googleapis.com/token",
+    };
+    const credentialsDir = mkdtempSync(path.join(os.tmpdir(), "vertex-adc-"));
+    const credentialsPath = path.join(credentialsDir, "adc.json");
+    writeFileSync(credentialsPath, JSON.stringify(credentials), {
+        mode: 0o600,
+        flag: "wx",
+    });
+    process.on("exit", () => {
+        rmSync(credentialsDir, { recursive: true, force: true });
+    });
+    process.env.GOOGLE_APPLICATION_CREDENTIALS = credentialsPath;
+    logger.info(`[RealtimeVoiceAgent] Vertex ADC written to ${credentialsPath} (project ${credentials.project_id}).`);
+}
+/**
+ * Force pure Vertex/ADC auth for the Gemini Live WebSocket.
+ *
+ * `@google/genai` 1.52+ uses a Gemini Developer API key for the realtime
+ * WebSocket auth even when `vertexai: true` and project/location are set, which
+ * Vertex rejects at the handshake (WS close 1006). The realtime worker only
+ * ever talks to Vertex, so remove these keys (only affects this process).
+ */
+export function clearGeminiApiKeyEnv() {
+    for (const key of ["GOOGLE_API_KEY", "GOOGLE_AI_API_KEY", "GEMINI_API_KEY"]) {
+        if (process.env[key]) {
+            delete process.env[key];
+            logger.info(`[RealtimeVoiceAgent] cleared ${key} so genai uses Vertex/ADC auth (not API key) for the Live WS.`);
+        }
+    }
+}

package/dist/voice/livekit/voiceAgent.js

@@ -192,9 +192,20 @@ export function defineVoiceAgent(config) {
     async function entry(ctx) {
         const entryStartedAt = Date.now();
         await ctx.connect();
-        logger.debug(`[LiveKitVoiceAgent] Joined room "${ctx.room.name}" in ${Date.now() - entryStartedAt}ms`);
-        // When the user actually stopped speaking (VAD), used to measure how long
-        // the agent waited after speech before committing the turn to the LLM.
+        logger.debug("voice.agent.roomJoined", {
+            room: ctx.room.name,
+            ms: Date.now() - entryStartedAt,
+        });
+        const { RoomEvent } = await import("@livekit/rtc-node");
+        ctx.room.on(RoomEvent.ParticipantDisconnected, () => {
+            if (ctx.room.remoteParticipants.size === 0) {
+                logger.info("voice.agent.participantLeft", {
+                    room: ctx.room.name,
+                    action: "shutdown",
+                });
+                ctx.shutdown("participant left");
+            }
+        });
         let userStoppedSpeakingAt;
         const neurolink = await config.createNeuroLink();
         const brain = createVoiceBrain({
@@ -245,11 +256,6 @@ export function defineVoiceAgent(config) {
                 final: false,
             });
         }
-        /**
-         * Lock the user bubble at turn-end and reset the buffer for the next turn.
-         * `replacesPrevious` tells the client this committed turn absorbed a prior
-         * interrupted turn, so it should remove the orphaned previous user bubble.
-         */
         function commitUserTranscript(finalText, replacesPrevious = false) {
             if (transcriptEmitter !== undefined) {
                 transcriptEmitter.emit("voice:user-transcript", {
@@ -274,7 +280,9 @@ export function defineVoiceAgent(config) {
                 pendingPrefix = "";
                 commitUserTranscript(promptText, hadPrefix);
                 if (userStoppedSpeakingAt !== undefined) {
-                    logger.debug(`[LiveKitVoiceAgent] Endpointing waited ${Date.now() - userStoppedSpeakingAt}ms before sending turn to LLM`);
+                    logger.debug("voice.agent.endpointingWaited", {
+                        ms: Date.now() - userStoppedSpeakingAt,
+                    });
                 }
                 return brainTurnStream(brain, promptText, conversationId, () => {
                     // Interrupted before producing any reply → carry this turn's text
@@ -299,7 +307,7 @@ export function defineVoiceAgent(config) {
         };
         if (eouTurnDetector !== undefined) {
             turnHandling.turnDetection = eouTurnDetector;
-            logger.info("[LiveKitVoiceAgent] Semantic end-of-utterance turn detection enabled (English)");
+            logger.info("voice.agent.eouEnabled", { language: "english" });
         }
         else if (config.turn?.mode) {
             turnHandling.turnDetection = config.turn.mode;
@@ -319,20 +327,11 @@ export function defineVoiceAgent(config) {
             tts,
             llm: new PlaceholderLLM(),
             turnHandling,
-            // Do NOT speculatively call the LLM on preflight transcripts before the
-            // turn ends — with NeuroLink as the brain each call is a real LLM request,
-            // and it makes the agent feel like it responds while you're still talking.
             preemptiveGeneration: false,
         });
         const agent = new NeuroLinkVoiceAgent({
             instructions: config.systemPrompt ?? "",
         });
-        // Inactivity watchdog: shut the per-call Job down after a stretch with no
-        // user or agent activity (mirrors Clairvoyance). On timeout `ctx.shutdown`
-        // runs the shutdown callbacks (disposing the bridge) and the Job process
-        // exits — freeing its RAM and the EOU model — while the browser observes a
-        // room disconnect. Reset on every interaction below. Configure via
-        // VOICE_INACTIVITY_TIMEOUT_MS (default 10 min); <= 0 disables the watchdog.
         const inactivityTimeoutMs = Number(process.env.VOICE_INACTIVITY_TIMEOUT_MS ?? 600_000);
         const inactivityEnabled = Number.isFinite(inactivityTimeoutMs) && inactivityTimeoutMs > 0;
         let inactivityTimer;
@@ -350,7 +349,11 @@ export function defineVoiceAgent(config) {
             clearInactivityTimer();
             inactivityTimer = setTimeout(() => {
                 inactivityFired = true;
-                logger.info(`[LiveKitVoiceAgent] Inactivity timeout (${Math.round(inactivityTimeoutMs / 1000)}s) reached — shutting down job for room "${ctx.room.name}"`);
+                logger.info("voice.agent.inactivityTimeout", {
+                    room: ctx.room.name,
+                    timeoutMs: inactivityTimeoutMs,
+                    action: "shutdown",
+                });
                 ctx.shutdown("inactivity timeout");
             }, inactivityTimeoutMs);
             // The watchdog must not, by itself, keep the event loop alive.
@@ -359,46 +362,53 @@ export function defineVoiceAgent(config) {
         ctx.addShutdownCallback(async () => {
             clearInactivityTimer();
         });
-        // Track when the user actually stops speaking (VAD) so endpointing latency
-        // can be measured, and reset the inactivity watchdog on user activity.
+        if (process.env.LK_REALTIME_CONNECT_MODE === "true") {
+            ctx.addShutdownCallback(async () => {
+                const parentPid = process.ppid;
+                setTimeout(() => {
+                    try {
+                        if (typeof parentPid === "number" && parentPid > 1) {
+                            process.kill(parentPid, "SIGTERM");
+                        }
+                    }
+                    catch {
+                        // Parent already gone — fall through to the hard exit below.
+                    }
+                    process.exit(0);
+                }, 500).unref?.();
+            });
+        }
         session.on(voice.AgentSessionEventTypes.UserStateChanged, (ev) => {
             noteActivity();
             if (ev.oldState === "speaking" && ev.newState !== "speaking") {
                 userStoppedSpeakingAt = Date.now();
             }
         });
-        // Reset the inactivity watchdog on any agent speech/processing and on every
-        // committed conversation item (user turn or agent reply), so the timeout
-        // only fires during a genuine lull in the conversation.
         session.on(voice.AgentSessionEventTypes.AgentStateChanged, () => {
             noteActivity();
         });
         session.on(voice.AgentSessionEventTypes.ConversationItemAdded, () => {
             noteActivity();
         });
-        // Forward user STT transcripts to the data-channel bridge as a single
-        // live-updating bubble. `UserInputTranscribed` fires `isFinal: true` per
-        // finalized SEGMENT (several per turn), so we never forward those as the
-        // turn-final; `emitUserTranscriptSegment` accumulates them into the per-turn
-        // buffer and emits `final: false`. The lone `final: true` is sent from
-        // `llmNode` at the real turn boundary.
         if (transcriptEventsEnabled) {
             session.on(voice.AgentSessionEventTypes.UserInputTranscribed, (ev) => {
                 emitUserTranscriptSegment(ev.transcript, ev.isFinal);
             });
         }
-        logger.info("[LiveKitVoiceAgent] Session starting", {
+        logger.info("voice.agent.sessionStarting", {
             room: ctx.room.name,
             provider,
             model,
         });
         await session.start({ agent, room: ctx.room });
-        // Start the inactivity countdown now that the session is live; every
-        // interaction handler above re-arms it.
+        if (config.greeting !== undefined && config.greeting.trim().length > 0) {
+            const greetingStream = brainTurnStream(brain, config.greeting, conversationId);
+            session.say(greetingStream, {
+                addToChatCtx: true,
+                allowInterruptions: true,
+            });
+        }
         noteActivity();
-        // Data-channel event bridge: forward NeuroLink events (text, tool calls,
-        // results, HITL prompts, status) to the browser, and accept HITL responses
-        // back. Only when enabled and the instance exposes its event emitter.
         if (config.events?.enabled === true && neurolink.getEventEmitter) {
             const bridge = await attachEventBridge({
                 room: ctx.room,

package/dist/voice/livekit/voiceAgentWorker.d.ts

@@ -11,6 +11,7 @@
  * See docs/features/livekit-voice-agent.md.
  */
 import type { LiveKitWorkerLaunchOptions } from "../../types/index.js";
+export declare function installVoiceWorkerProcessGuards(metricsIntervalMs?: number): void;
 /**
  * Launch the LiveKit voice agent worker.
  *
@@ -25,3 +26,4 @@ import type { LiveKitWorkerLaunchOptions } from "../../types/index.js";
  * ```
  */
 export declare function startVoiceAgentWorker(options: LiveKitWorkerLaunchOptions): Promise<void>;
+export declare function startRealtimeVoiceAgentWorker(options: LiveKitWorkerLaunchOptions): Promise<void>;

package/dist/voice/livekit/voiceAgentWorker.js

@@ -11,8 +11,58 @@
  * See docs/features/livekit-voice-agent.md.
  */
 import { resolveEouTurnDetection, resolveLiveKitServerConfig, } from "./config.js";
+import { logger } from "../../utils/logger.js";
 const DEFAULT_AGENT_NAME = "neurolink-voice";
 const EOU_METHOD_MULTILINGUAL = "lk_end_of_utterance_multilingual";
+const IS_JOB_CHILD = process.argv.some((arg) => arg.includes("job_proc"));
+const PROC_ROLE = IS_JOB_CHILD ? "job(child)" : "worker(parent)";
+let processGuardsInstalled = false;
+export function installVoiceWorkerProcessGuards(metricsIntervalMs = Number(process.env.VOICE_METRICS_INTERVAL_MS ?? 10000)) {
+    if (processGuardsInstalled) {
+        return;
+    }
+    processGuardsInstalled = true;
+    const procInfo = {
+        role: PROC_ROLE,
+        pid: process.pid,
+        ppid: process.ppid,
+    };
+    process.on("uncaughtException", (error) => {
+        logger.error("voiceWorker.uncaughtException", {
+            ...procInfo,
+            error: error?.stack ?? String(error),
+        });
+        if (IS_JOB_CHILD) {
+            setTimeout(() => process.exit(1), 1000).unref?.();
+        }
+    });
+    process.on("unhandledRejection", (reason) => {
+        logger.error("voiceWorker.unhandledRejection", {
+            ...procInfo,
+            error: reason instanceof Error ? reason.stack : String(reason),
+        });
+    });
+    for (const signal of ["SIGTERM", "SIGINT", "SIGHUP"]) {
+        process.on(signal, () => {
+            logger.warn("voiceWorker.signal", { ...procInfo, signal });
+            setTimeout(() => process.exit(0), 1500);
+        });
+    }
+    if (Number.isFinite(metricsIntervalMs) && metricsIntervalMs > 0) {
+        const mb = (bytes) => Math.round((bytes / 1024 / 1024) * 10) / 10;
+        const timer = setInterval(() => {
+            const usage = process.memoryUsage();
+            logger.debug("voiceWorker.mem", {
+                ...procInfo,
+                rssMb: mb(usage.rss),
+                heapUsedMb: mb(usage.heapUsed),
+                heapTotalMb: mb(usage.heapTotal),
+                externalMb: mb(usage.external),
+            });
+        }, metricsIntervalMs);
+        timer.unref?.();
+    }
+}
 /**
  * Register the English EOU inference runner in the worker process.
  *
@@ -55,3 +105,17 @@ export async function startVoiceAgentWorker(options) {
         apiSecret: server.apiSecret,
     }));
 }
+export async function startRealtimeVoiceAgentWorker(options) {
+    installVoiceWorkerProcessGuards();
+    if (process.env.LIVEKIT_EOU_TURN_DETECTION) {
+        delete process.env.LIVEKIT_EOU_TURN_DETECTION;
+        logger.info("realtime.worker.eouDisabled", {
+            reason: "s2s-in-model-turn-detection",
+        });
+    }
+    if (process.argv.includes("connect")) {
+        process.env.LK_REALTIME_CONNECT_MODE = "true";
+        logger.info("realtime.worker.connectMode", { enabled: true });
+    }
+    await startVoiceAgentWorker(options);
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@juspay/neurolink",
-  "version": "9.71.0",
+  "version": "9.73.0",
   "packageManager": "pnpm@10.15.1",
   "description": "Universal AI Development Platform with working MCP integration, multi-provider support, voice (TTS/STT/realtime), and professional CLI. 58+ external MCP servers discoverable, multimodal file processing, RAG pipelines. Build, test, and deploy AI applications with 21+ providers: OpenAI, Anthropic, Google AI Studio, Google Vertex, AWS Bedrock, Azure OpenAI, Mistral, LiteLLM, SageMaker, Hugging Face, Ollama, OpenAI-compatible, OpenRouter, DeepSeek, NVIDIA NIM, LM Studio, llama.cpp, plus voice (OpenAI TTS, ElevenLabs, Deepgram, Azure Speech).",
   "author": {
@@ -387,6 +387,7 @@
     "@livekit/agents-plugin-cartesia": "^1.4.5",
     "@livekit/agents-plugin-deepgram": "^1.4.5",
     "@livekit/agents-plugin-elevenlabs": "^1.4.5",
+    "@livekit/agents-plugin-google": "^1.4.5",
     "@livekit/agents-plugin-livekit": "^1.4.5",
     "@livekit/agents-plugin-silero": "^1.4.5",
     "@livekit/agents-plugin-soniox": "^1.4.5",