@byokit/realtime 0.3.0 → 0.3.2

package/CHANGELOG.md

@@ -2,6 +2,16 @@
 
 ## Unreleased
 
+## 0.3.2 (2026-10-17)
+
+- Dependency update: ws 8.22.0.
+
+## 0.3.1 (2026-10-01)
+
+
+
+- SECURITY: Lazy WebRTC capture no longer opens the Android microphone or sends untapped speech before `attachMic()`; `releaseMic()` closes recording while keeping the call connected. Apps using lazy capture should update.
+
 ## 0.3.0 (2026-10-01)
 
 

package/README.md

@@ -194,8 +194,10 @@ function endCall() { client.stop(); }
 ```
 
 `attachMic()` awaits the app's microphone permission/service lease, captures
-an audio track and uses `sender.replaceTrack(track)`. `releaseMic()` awaits
-`replaceTrack(null)`, stops capture tracks and releases the lease so the app
+an audio track and uses `sender.replaceTrack(track)` before enabling the sender's
+encoding. Lazy peers negotiate `sendrecv` with an inactive encoding: Android's
+audio device cannot start recording just because an answer is applied.
+`releaseMic()` disables that encoding before `replaceTrack(null)`, stops capture tracks and releases the lease so the app
 can stop its microphone service and clear the OS indicator. Neither operation
 renegotiates. Repeated calls are safe; release or stop during pending capture
 cleans up when that capture resolves. Reconnecting starts another lazy peer;
@@ -256,4 +258,4 @@ Credentials, endpoint restrictions and redirect rejection are unchanged.
 The identity option allows a consumer to compare signaling latency; offline
 tests prove startup ordering, not a reduction in the remote provider's POST time.
 
-For an explicitly live microphone and subscription-provider proof, run [the bounded voice demo](../../examples/realtime-voice/README.md). It signs in through accounts, keeps credentials only in memory, and records connection facts and turn counts. Offline tests do not establish a live pass.
+For an explicitly live microphone and subscription-provider proof, run [the bounded voice demo](../../examples/realtime-voice/README.md). It signs in through accounts, keeping the sign-in at this computer's machine store, and records connection facts and turn counts. Offline tests do not establish a live pass.

package/dist/chatgpt.js

@@ -1,5 +1,5 @@
-import { cleanProse, cleanProseBytes } from "./prose.js";
-import { MAX_REALTIME_SDP_BYTES } from "./frames.js";
+import { cleanProse, cleanProseBytes } from './prose.js';
+import { MAX_REALTIME_SDP_BYTES } from './frames.js';
 export const CHATGPT_SIGNALING_URL = 'https://chatgpt.com/backend-api/codex/realtime/calls?intent=quicksilver&architecture=avas';
 function emit(frame) { process.stdout.write(`${JSON.stringify(frame)}\n`); }
 export async function boundedBody(response, max = MAX_REALTIME_SDP_BYTES) {

package/dist/child.js

@@ -8,8 +8,8 @@ var __rewriteRelativeImportExtension = (this && this.__rewriteRelativeImportExte
 };
 import { createInterface } from 'node:readline';
 import { randomUUID } from 'node:crypto';
-import { parseRealtimeClientFrame } from "./frames.js";
-import { createChatgptAdapter } from "./chatgpt.js";
+import { parseRealtimeClientFrame } from './frames.js';
+import { createChatgptAdapter } from './chatgpt.js';
 const input = createInterface({ input: process.stdin, crlfDelay: Infinity });
 let adapter;
 const pending = new Map();

package/dist/client.js

@@ -1,4 +1,4 @@
-import { parseRealtimeClientFrame, parseRealtimeHostFrame, realtimePcm16ByteLength } from "./frames.js";
+import { parseRealtimeClientFrame, parseRealtimeHostFrame, realtimePcm16ByteLength } from './frames.js';
 /** Provider-blind client with bounded reconnects and exclusive media ownership. */
 export function realtimeClient(options) {
     const audio = options.audio;

package/dist/index.js

@@ -1,7 +1,7 @@
-export * from "./frames.js";
-export * from "./types.js";
-export * from "./words.js";
-export * from "./prose.js";
-export * from "./client.js";
-export * from "./auth.js";
-export { providers } from "./catalogue.js";
+export * from './frames.js';
+export * from './types.js';
+export * from './words.js';
+export * from './prose.js';
+export * from './client.js';
+export * from './auth.js';
+export { providers } from './catalogue.js';

package/dist/node.js

@@ -1,12 +1,12 @@
 import { spawn } from 'node:child_process';
 import { createInterface } from 'node:readline';
 import { fileURLToPath } from 'node:url';
-import { parseRealtimeClientFrame, parseRealtimeHostFrame } from "./frames.js";
-import { RealtimeError } from "./types.js";
-import { cleanProseBytes } from "./prose.js";
-import { realtimeAuthCheck } from "./auth.js";
-export { toolBridge, appBridge, delegationHandler } from "./tools.js";
-export { realtimeAuthCheck } from "./auth.js";
+import { parseRealtimeClientFrame, parseRealtimeHostFrame } from './frames.js';
+import { RealtimeError } from './types.js';
+import { cleanProseBytes } from './prose.js';
+import { realtimeAuthCheck } from './auth.js';
+export { toolBridge, appBridge, delegationHandler } from './tools.js';
+export { realtimeAuthCheck } from './auth.js';
 /** Provider sockets and SDP signaling run in a kit-owned child, never in the app process. */
 export function realtimeEngine(options) {
     if (!['chatgpt', 'openai', 'gemini', 'xai'].includes(options.engine))

package/dist/providers/gemini.js

@@ -1,6 +1,6 @@
 import WebSocket from 'ws';
-import { responseFence } from "../policy.js";
-import { cleanProse, providerRefusal as refusal } from "../prose.js";
+import { responseFence } from '../policy.js';
+import { cleanProse, providerRefusal as refusal } from '../prose.js';
 // Lifted wire behavior; this module is loaded only by the kit child process.
 export function createAdapter(options) {
     const MODEL = options.model ?? 'gemini-3.8-live';

package/dist/providers/openai.js

@@ -1,6 +1,6 @@
 import WebSocket from 'ws';
-import { responseFence } from "../policy.js";
-import { cleanProse, providerRefusal as refusal } from "../prose.js";
+import { responseFence } from '../policy.js';
+import { cleanProse, providerRefusal as refusal } from '../prose.js';
 // Lifted wire behavior; this module is loaded only by the kit child process.
 export function createAdapter(options) {
     const MODEL = options.model ?? 'gpt-realtime-2.1';

package/dist/providers/xai.js

@@ -1,6 +1,6 @@
 import WebSocket from 'ws';
-import { responseFence } from "../policy.js";
-import { cleanProse, providerRefusal as refusal } from "../prose.js";
+import { responseFence } from '../policy.js';
+import { cleanProse, providerRefusal as refusal } from '../prose.js';
 // Lifted wire behavior; this module is loaded only by the kit child process.
 export function createAdapter(options) {
     const MODEL = options.model ?? 'grok-voice-think-fast-2.0';

package/dist/rn-webrtc.js

@@ -1,4 +1,4 @@
-import { webRtcPeer as createPeer } from "./webrtc.js";
+import { webRtcPeer as createPeer } from './webrtc.js';
 /** Native media stays unloaded until the app actually starts a voice session. */
 export async function webRtcPeer(options) {
     if (options.platform)

package/dist/tools.js

@@ -1,5 +1,5 @@
 import { randomUUID } from 'node:crypto';
-import { cleanProseBytes } from "./prose.js";
+import { cleanProseBytes } from './prose.js';
 /** Bounded, deduplicated tool execution. Product authorization stays in handlers. */
 export function toolBridge(options) {
     const lifetime = new AbortController();

package/dist/webrtc.js

@@ -1,4 +1,4 @@
-import { MAX_REALTIME_SDP_BYTES, MAX_REALTIME_WEBRTC_DATA_BYTES } from "./frames.js";
+import { MAX_REALTIME_SDP_BYTES, MAX_REALTIME_WEBRTC_DATA_BYTES } from './frames.js';
 function sdp(value) {
     if (typeof value !== 'string' || !value.startsWith('v=0') || value.includes('\0') || new TextEncoder().encode(value).length > MAX_REALTIME_SDP_BYTES)
         throw new Error('Invalid voice connection answer.');
@@ -47,6 +47,19 @@ export async function webRtcPeer(options) {
         options.onError(error);
         stop();
     } };
+    const setSending = async (active) => {
+        if (options.capture !== 'lazy')
+            return;
+        const parameters = sender.getParameters();
+        if (!parameters.encodings.length)
+            throw new Error('Voice capture control is unavailable.');
+        parameters.encodings.forEach(encoding => { encoding.active = active; });
+        await sender.setParameters(parameters);
+        // The native bridge can resolve even when the engine rejects setParameters.
+        const applied = sender.getParameters().encodings;
+        if (!applied.length || applied.some(encoding => encoding.active !== active))
+            throw new Error('Voice capture control failed.');
+    };
     const captureMic = async (generation) => {
         await options.audio.microphone.acquire();
         acquired = true;
@@ -79,14 +92,28 @@ export async function webRtcPeer(options) {
                 const captured = await captureMic(generation);
                 if (captured) {
                     await sender.replaceTrack(captured.getAudioTracks()[0]);
+                    if (!stopped && generation === micEpoch)
+                        await setSending(true);
                     if (stopped || generation !== micEpoch) {
-                        if (!stopped)
+                        if (!stopped) {
+                            await setSending(false);
                             await sender.replaceTrack(null);
+                        }
                         clearMic();
                     }
                 }
             }
             catch (error) {
+                // Never leave a trackless native send stream recording after a failure.
+                try {
+                    if (!stopped && options.capture === 'lazy') {
+                        await setSending(false);
+                        await sender.replaceTrack(null);
+                    }
+                }
+                catch {
+                    stop();
+                }
                 clearMic();
                 throw error;
             }
@@ -97,8 +124,15 @@ export async function webRtcPeer(options) {
         stream?.getAudioTracks().forEach(track => { track.enabled = false; });
         return serializeMic(async () => {
             try {
-                if (!stopped && stream)
+                if (!stopped && stream) {
+                    await setSending(false);
                     await sender.replaceTrack(null);
+                }
+            }
+            catch (error) {
+                if (options.capture === 'lazy')
+                    stop();
+                throw error;
             }
             finally {
                 clearMic();
@@ -153,8 +187,12 @@ export async function webRtcPeer(options) {
             options.onConnectionState('connecting'); };
         current.ontrack = event => { const track = event.track; if (track.kind !== 'audio' || stopped)
             return; options.onRemoteAudio(!track.muted); track.onmute = () => options.onRemoteAudio(false); track.onunmute = () => options.onRemoteAudio(true); track.onended = () => options.onRemoteAudio(false); };
-        if (options.capture === 'lazy')
-            sender = current.addTransceiver('audio', { direction: 'sendrecv' }).sender;
+        if (options.capture === 'lazy') {
+            // Android starts the device module for an active send stream even without
+            // a track. Disable the encoding before either description can be applied.
+            sender = current.addTransceiver('audio', { direction: 'sendrecv', sendEncodings: [{ active: false }] }).sender;
+            await setSending(false);
+        }
         else {
             const input = stream.getAudioTracks()[0];
             if (!input)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@byokit/realtime",
-  "version": "0.3.0",
+  "version": "0.3.2",
   "description": "Realtime voice sessions with host-held credentials and injected audio.",
   "license": "Apache-2.0",
   "type": "module",
@@ -43,7 +43,7 @@
     "access": "public"
   },
   "dependencies": {
-    "ws": "8.21.3"
+    "ws": "8.22.0"
   },
   "peerDependencies": {
     "react-native-webrtc": "124.0.8"