Skip to content

Limits and support

Where the SDK runs, what has been exercised and where, and the limits it has today.

Package Runs on Checked in CI
reactor-effect-client Node 22 or newer, Bun, browsers Its suites on Node 22.22 and 24.15 (Ubuntu 24.04), Node 24.15 (macOS 15), and Bun 1.4.2
reactor-effect-browser Browsers with WebRTC and Web Crypto, on localhost or HTTPS Its suites against DOM fakes on Node and Bun; a browser bundle that must reach no Node code
reactor-effect-native Node 22 or newer and Bun, on Linux x64 (glibc) and macOS arm64 Rust checks and the native suites on Node and Bun, on Ubuntu 24.04 and macOS 15, against the built addon
  • The isolated native host, NativePeer.layerIsolated(), needs a Node parent process. Under Bun it fails to build with UnsupportedCapability.
  • The macOS addon targets macOS 13.0 or later.
  • Real Chrome exchanging media with the native host runs in CI’s browser integration test, on Node.

On hosted Reactor (every run), in paid runs from 2026-09-24 to 2026-09-29 on the 0.3.0-rc.0 to 0.8.0 libraries: sessions, tokens and their refresh, adoption of a dead owner’s session, reconnects, uploads and references, H3’s queue commands, moderation, and Playout over three sessions, with one planned renewal and one replacement of an ended session. They ran from Linux x64, through the native host in process under Bun and isolated under Node, over a direct network path. On 0.9.0 only the one-session showreel check (2026-10-01) has run; 0.9.0’s other checks, its published client packages, the browser host and the in-process native host under Node have not run on hosted Reactor yet. Vidu S2-Avatar’s avatar, calls, picture, voice and refusals ran through the raw Session in two avatar probes (2026-10-01 and 10-02), and its provider in two character runs on Effect 4.0.0 (2026-10-02): one passed every criterion, and in the other the native connection failed with a Protocol failure right after the provider’s first command, for a cause not yet known.

FastH3 has one paid raw Session probe (2026-10-03). It measured sample lengths and read the state, queues and history; it refused the undeclared reference_images: [] key. This probe did not use the FastH3 provider or playout source; neither has been qualified on hosted Reactor.

Locally and in CI only:

  • the browser host, against DOM fakes and, for media, real Chrome talking to the native host;
  • the macOS addon;
  • a TURN relay, once, locally: real Chrome and the native bridge through a loopback Coturn server on 2026-09-22, before the current API;
  • the native media path under load, against a local libwebrtc sender: see measured performance;
  • 0.9.0’s Playout.place and follows, on ReactorTest;
  • the ViduS2Avatar provider and the avatar example, on ReactorTest;
  • the FastH3 provider, on ReactorTest.

Not yet exercised anywhere: TURN relays with hosted credentials, native media published on to a browser or a stream ingest, sessions longer than a few minutes on hosted Reactor, and dedicated physical machines: the native performance figures come from a 4-vCPU container and GitHub’s runners.

ReactorTest models H3, FastH3 and Vidu S2-Avatar. H3’s and Vidu’s delays are drawn from paid runs; FastH3 borrows H3’s timings, and its fastH3History defaults to eight because hosted history capacity remains unmeasured. It does not model hosted latency as a real network varies it, billing as Reactor charges it, TURN relays or interop. A rehearsal on it is never evidence of hosted behaviour.

The SDK targets H3 Reference Turbo Realtime (reactor/h3-reference-to-video-turbo-realtime), against its documented 0.5.5 schema, FastH3 (reactor/fast-h3) through FastH3, and Vidu S2-Avatar (reactor/vidu-s2-avatar) through ViduS2Avatar. H3.make reads the deployment’s schema and fails any command the deployment lacks with UnsupportedCapability before sending it; the hosted deployment reports its version as v0.0.0.

Reactor, Session and the hosts are not specific to H3: reactor.create takes any model name, and session.command(name, data) sends any command. H3, FastH3 and Vidu S2-Avatar have providers; H3 and FastH3 have playout sources. Create each session with its provider’s modelName. H3.make checks declared commands rather than identifying the model, so it can admit a FastH3 deployment; use FastH3.make to send FastH3’s frame arguments. FastH3.make refuses an H3 deployment with UnsupportedCapability before sending model commands. On Node and Bun, a Vidu S2-Avatar caller speaks only through say: the native host can’t publish a microphone track.

  • One incoming video track and one incoming audio track. The pinned libwebrtc binding reports a received track without the identity needed to match two tracks of one kind to the negotiated ones, so a session that declares more fails with UnsupportedCapability before negotiation. H3 sends one of each. Outgoing tracks may be several.
  • Bounded queues. The addon holds at most 8 frames, 256 PCM blocks and 1,024 events, and each reader at most 24 frames or 128 PCM blocks. A full media queue drops its oldest item and counts it; a full event queue ends the connection with Overflow; a reader that falls further behind fails alone with Overflow.
  • One libwebrtc factory per process. If a connection’s shutdown outlives shutdownTimeout in process, no new peer is made until it completes. The isolated host kills the child instead.
  • Decoding failures are not reported, and neither is ICE’s connection state: the binding surfaces neither. A failed connection is classified from its statistics as IceFailed or TransportFailed.
  • It has platform tracks only, no decoded frames: session.decoded and a playout’s video and audio fail with UnsupportedCapability.
  • A message larger than the data channel’s negotiated bound (at most 256 KiB) fails the connection with Overflow, and a send is refused while more than 1 MiB waits in the channel’s buffer.
  • Effect ~4.0.0: 4.0.x patches, not 4.1. The Effect modules the SDK builds on, such as effect/http and effect/rpc, are marked unstable, and Effect may change those in a minor release, so Effect 4.1 needs a new SDK release. Projects that install @effect/platform-node also pin @effect/platform-node-shared; see Installation.
  • 0.x. The three packages are released together under one version, and while it is 0.x a minor version may break the API: 0.8.0 and 0.9.0 both did. Pin exact versions. Each release’s CHANGELOG.md entry has an upgrade table.
  • Declarations without DOM types. Effect 4.0.0 names the DOM’s TextDecoderOptions in its declarations; a Node project that omits DOM types and checks library declarations declares that interface itself.

A request H3 would refuse fails locally with outcome not-submitted, before anything is uploaded:

  • clip length 5 to 15.084 seconds, aligned up to H3’s frame grid (124 frames at 24 fps, then steps of 17);
  • up to nine reference images, and up to three reference audios (two when the clip continues another), twelve references in all; audio needs an image or a continuation;
  • metadata up to 2,000 characters.

H3’s text budget, about 2,000 tokens, is enforced by H3 at build: a longer prompt is accepted and then fails its clip. See the H3 provider.

FastH3.Request accepts 5.167–14.375 s. Lengths align up on its grid of 124 frames, then steps of 17, at 24 fps: the rounded 5.167 s minimum represents 124 frames, and 6 s builds 158 frames (6.583 s). start and end each take an image reference or a clip id, for an opener and closer; there are no H3 reference-image or audio lists. Metadata shares the 2,000-character bound, including the acceptance envelope. Reactor enforces the 1,024-token prompt budget at enqueue. The paid probe sampled this length rule; it did not measure every length, history capacity, dependent build order or the documented stop drain. See FastH3.

FastH3Source.model plans a playout item without seconds at 5.167 s; calling the provider or source directly preserves the session’s length setting, initially 14.375 s. Automatic continuity uses only built playing, playout or history clips on the same session, excluding failed clips and generation entries. An explicit start wins. A missing hint is dropped before the single send with reactor.continuation: "dropped"; post-send failures and dispatch outcomes pass through unchanged, with no failure-triggered refresh or resend. See Air FastH3 through the playout.

Credentials stay Redacted from configuration to the request that uses them; provider text stays out of messages, logs and spans; native queues are bounded; and the client and browser packages never load native code. Report a vulnerability privately, as SECURITY.md describes.