LiveKit

Recently shipped Egress v1.14.0 with a rewritten A/V sync engine and an Egress v2 API, plus new TTS plugins for agents.

Egress v1.14.0 rewrote A/V synchronization around per-track drift probes and audio tempo control. The new sync engine and two-tiered pacer replace the old A/V lockstep, and the Egress v2 API implementation lands alongside sizeable resilience work:

• Egress now drains on sustained high CPU instead of hard-killing, uploading what was recorded.
• A dedicated monitor goroutine reaps orphaned PulseAudio sinks; wedged RTMP sinks are detected and isolated per stream, so multi-output egresses keep streaming on surviving outputs.
• Missing keyframes re-request PLI across video codecs; upload metrics report error status class (4xx/5xx/internal).
• Chrome 150 support.

Rust SDKs introduced pluggable signalling transport via the new livekit-net crate. The signalling WebSocket and pre-connect HTTP calls route through pluggable traits (WsClient, HttpClient) from a process-global registry. This is a breaking change: SignalError::WsError is removed, TLS handshake failures surface HTTP status as SignalError::Client/Server, and transport failures use the new SignalError::Connection/Closed variants. The compiler caught an RPC server version parsing typo.

Rust SDKs fixed a security bug in NativeVideoSource keepalive frames. webrtc::I420Buffer::Create left pixel planes uninitialized, so pre-capture keepalive frames could leak recycled heap contents to subscribers as first keyframes — now they emit black frames. NVIDIA NVENC AV1 encoding is available when the GPU supports it, and data tracks forward per-schema metadata.

Agents gained two new TTS plugins — Deepgram Flux (v2/speak) and Bland — and now forward agent_false_interruption over RemoteSession. The Phonic plugin exposes more realtime config options (VAD tuning, backchannels, redaction, MCP servers), and each TTS fallback recovery path gets its own task slot. JS core bumped to 1.6.2.

Server v1.13.4–v1.13.5 and CLI v2.18.0–v2.18.2 shipped the bulk of the remaining polish: SIP dialing mocks, IPv6 candidate controls, more h264 profiles, a drop-on-close test scenario, and a redaction flag in Job; the CLI added deployment support, --audio simulation mode, attribute tool-call attribution in the TUI transcript, and migrated from the Docker client to moby.