The WebRTC pipeline

webrtc/WebRtcClient.kt (~1,500 lines) is the entire media engine. It owns every native WebRTC object and exposes a plain-Kotlin API to MainActivity.

Construction


WebRtcClient(
    context: Context,
    eglBase: EglBase,                       // shared GL context for capture + render
    micMode: MicMode = MicMode.DEFAULT,
    onIceCandidateGenerated: (IceCandidate) -> Unit,
    onConnectionStateChanged: (PeerConnection.PeerConnectionState) -> Unit,
)

MainActivity builds one after camera + mic permission is granted, and rebuilds it when a call ends (webRtcClient = null then initWebRtcClient()), or when a mic-mode change requires a hardware-stack rebuild.

What it owns

Field	Purpose
`peerConnectionFactory`	Built once with the encoder/decoder factories and the ADM
`audioDeviceModule`	`JavaAudioDeviceModule` with HW AEC/NS toggled per mic mode and an external processing chain
`peerConnection`	The single `PeerConnection` (`UNIFIED_PLAN`, `GATHER_CONTINUALLY`)
`videoCapturer`	`CameraVideoCapturer` - Camera2 if supported, else Camera1
`videoSource` / `localVideoTrack`	`isScreencast = false` so WebRTC adapts resolution down under load and keeps FPS smooth. Screen sharing currently reuses this source (a dedicated screencast source is a pending improvement)
`screenCapturer`	`ScreenCapturerAndroid` feeding the same `VideoSource` via its `capturerObserver`
`audioSource` / `localAudioTrack`	Opus source with per-mode constraints
`videoSender` / `audioSender`	`RtpSender`s - hot-swap tracks with `setTrack(track, false)` (no renegotiation)
`remoteVideoTrack` / `remoteAudioTrack`	Set in `onTrack`
`surfaceTextureHelper` / `screenTextureHelper`	Separate helpers - sharing one leaves the MediaProjection virtual display with an invalid producer surface and zero frames delivered
`zoomController`	`CameraZoomController` - Camera2 zoom ratio, or digital crop fallback
RNNoise / gain / level probes	The external audio processors, chained in `CompositeAudioProcessing`

ICE / RTC configuration


val iceServers = listOf(
    IceServer.builder("stun:stun.l.google.com:19302").createIceServer(),
    IceServer.builder("stun:stun1.l.google.com:19302").createIceServer(),
)
RTCConfiguration(iceServers).apply {
    sdpSemantics = UNIFIED_PLAN
    continualGatheringPolicy = GATHER_CONTINUALLY
}

STUN only, Google’s public servers. There is no TURN - a call between two symmetric-NAT peers with no shared LAN will not connect. This is a known limitation (see roadmap).
GATHER_CONTINUALLY so a network change (Wi-Fi ↔ cellular) surfaces new candidates without a full ICE restart where possible.

Public API (1:1 today)

Method	Effect
`createPeerConnection()`	Build the PC, add local tracks, apply bitrate + codec prefs
`createOffer(cb)`	`createOffer` → munge SDP → `setLocalDescription` → `cb(sdp)`
`handleRemoteOffer(sdp, cb)`	`setRemoteDescription` → `createAnswer` → munge → `setLocalDescription` → `cb(answer)`
`handleRemoteAnswer(sdp)`	`setRemoteDescription`
`addIceCandidate(c)`	Buffered until the remote description is set, then drained
`restartIce()`	`createOffer` with `iceRestart = true`
`attachRemoteRenderer(sink)` / `detachRemoteRenderer`	Wire a Compose renderer to `remoteVideoTrack`
`attachLocalRenderer(sink)` / `detachLocalRenderer`	…to `localVideoTrack`; also kicks `ensureCapturing()`
`setLocalAudioEnabled(b)` / `setLocalVideoEnabled(b)`	Mute / camera-off (fully stops the capturer when off)
`switchCamera(cb)` / `setCameraFacing(front, cb)`	Flip
`setVideoQuality(w, h, fps, maxKbps, lockRes)`	Change capture format + encoder cap
`setPreferredCodec(name)`	Re-order codec preferences (`AV1` / `H265` / `H264`)
`setMicMode(mode)`	Live-swap the audio track / software APM; defer HW rebuild
`setMicGain(g)` / `setMicOverboost(b)`	Output gain on captured PCM
`startScreenShare(projection, w, h, fps, maxKbps, onStop)` / `stopScreenShare()`	Screencast
`startStatsPolling()` / `stopStatsPolling()`	The debug-panel loop - off by default
`setZoom(ratio)`	Pinch-zoom
`pauseCapture()` / `ensureCapturing()`	Lifecycle-driven camera control
`close()`	`@Synchronized` full teardown

SDP munging

Both createOffer and createAnswer run their local description through withRaisedLevel() before setLocalDescription:


private fun withRaisedLevel(desc: SessionDescription) =
    SessionDescription(desc.type, tuneOpus(raiseH265Level(desc.description)))

raiseH265Level rewrites every H.265 a=fmtp line to level-id=153 (Level 5.1, a 4K@60 budget). The SDK advertises level-id=93 (Level 3.1, ~720p), and strict hardware decoders reject frames above the negotiated level.
tuneOpus sets stereo=1 + sprop-stereo=1, maxaveragebitrate to 256 kbps, useinbandfec=1, and disables DTX.

Full detail in media negotiation.

Stats

pollStats() runs on a coroutine only while startStatsPolling() is active (panel visible + Activity STARTED). It calls peerConnection.getStats, and parseStats diffs bytesSent / bytesReceived against the previous poll to approximate bitrate, reads resolution / FPS / codec / RTT / loss from the report, and merges in a DeviceLoadSampler reading (CPU %, battery °C, watts). Output is a CallStats snapshot. See stats & device load.