Wire protocol

The wire protocol is deliberately small: JSON packets for everything addressable, plus a binary framing for stream payloads. It rides on a WebSocket, an HTTP request/response pair, Server-Sent Events, a MessagePort to a worker (a Service Worker or a SharedWorker), a WebRTC data channel, a WebTransport session or a message broker — the packets are identical on every one of them, which is what lets the same client code work over any, and behind a worker.

Stability

This page is the wire protocol at revision 2 — what @alexify/wrpc@2.x speaks, next to revision 1, which it still speaks to a 1.0 peer (Versioning). An independent implementation written against it keeps working for the life of the major version. It has three tiers, and the promise differs by tier:

• The core — the packets, their fields and their meanings, the error codes, batch frames, binary chunks, introspection, sessions and reconnect — does not change inside 2.x. A new optional field may be added; an existing one will not change shape or disappear. Unknown packet types are answered with a callback carrying code 500 and unknown fields are ignored, which is what makes an additive change safe for an older peer; error codes keep their meanings, and new codes may appear for new failure modes. Revision 2 changed one thing in the core 1.0 shipped: a binary frame whose first byte is 0x00 is a framed message, and a packet holding bytes travels as one. No JSON packet changed — and because a 1.0 peer reads no framed message, that one thing is what the revision marker negotiates.
• The carrier conventions — how a transport carries the core: connection metadata and its carriers, the enc negotiation on ping/pong, the SSE channel handshake, the rooms envelope, the WebRTC framing and signaling messages, the compression rule. Additive within 2.x under the same rule as the core. Most of them are new since 1.0 — every ## section 1.0 did not have carries a since 2.0 — and none of them asks anything of a 1.0 client or server: the table under Changes since 1.0 says what happens when the versions meet. What is left to configure is between the instances of one fleet, which have no handshake to negotiate on.
• The experimental sections — WebTransport, the broker binding, session encryption with sealed requests and broker sealing — describe revision 1 of a carrier or a format that may change in a minor, with the change in the CHANGELOG; each says so at its top, and its subpath is marked @experimental (stability).

The JavaScript API on top of this is versioned by the package's own semver and is a separate promise from the wire format. A change to a stable tier of this page is a major version, with the reasoning in the CHANGELOG.

Versioning

The wire carries a revision marker. Revision 1 (wrpc.v1) is what 1.0 speaks. Revision 2 (wrpc.v2) is revision 1 plus the framed messages: a packet whose byte values travel as bytes. A 2.x peer speaks both, and the rule is one line — a framed message is sent only where the peer said it reads one. Everywhere else a packet holding bytes travels as the JSON 1.0 made of it ({ "0": 137, … } for a typed array, { "type": "Buffer", "data": [ … ] } for a Node Buffer), so a 1.0 client and a 1.0 server keep working against 2.x with nothing to set.

On a WebSocket the revision is negotiated as a subprotocol:

Sec-WebSocket-Protocol: wrpc.v2, wrpc.v1   (client offer, newest first)
Sec-WebSocket-Protocol: wrpc.v2            (server selection)

A wRPC client offers the revisions it speaks; a wRPC server with no app-configured protocols/handleProtocols selects the newest one offered, wherever it sits in the list. Both sides therefore know, before the first packet, which revision the connection speaks — the selected name is on connection.protocol server-side, and the number on client.revision at both ends.

The rules that keep this compatible in every direction:

• A 1.0 client offers wrpc.v1 alone, and gets it. A 1.0 server picks wrpc.v1 out of a 2.x client's offer. Either way the connection speaks revision 1.
• A peer that offers nothing gets no subprotocol and the connection speaks revision 1 — the pre-marker handshake stays valid forever.
• An end that sends and reads no framed messages — attachments: false, or an injected packet codec — has nothing of revision 2 to offer: it offers, or selects, wrpc.v1 alone. Two 2.x ends whose options disagree therefore settle on revision 1 instead of one refusing the other's frames. The price is that a wrpc.v1 selection is ambiguous: it is what a 1.0 server answers and what a 2.x server without frames answers, and nothing on the handshake tells them apart. It matters only for what a 1.0 server does not read — a browser's carrier tokens — and there a client chooses its carrier rather than relying on the guess.
• A server whose app configures its own protocols list takes over negotiation entirely; the connection speaks revision 2 only when wrpc.v2 is what was selected.
• A future wrpc.v3 will be offered ALONGSIDE the older names, so an old server picks the one it knows and nothing breaks.
• The name wrpc. is reserved as a prefix: applications must not mint their own subprotocols under it. Besides revisions it holds the carrier tokens of connection metadata — offers that are read, never selected.

The other carriers have no subprotocol, and each says the revision where it can. Between two 2.x ends, the four that negotiate it look like this:

A 1.0 peer on the other end answers wrpc.v1, wrpc-version: 1 or a plain pong, and the connection speaks revision 1 from then on.

Carrier	How the revision is said
WebSocket	the subprotocol, above
HTTP	the response header wrpc-version, and the request's Accept
SSE	nothing to say — always revision 1: the transport is text-only, carries no framed message, and a packet's bytes travel as JSON both ways
Worker port	v on the port's first ping and its pong
WebTransport	f in the capabilities message each end sends first (below)
WebRTC	f in a description signal's caps (below), per half of the peer; over a raw channel, the applications' agreement
The broker binding	wrpc-version on a session's hello and welcome (the binding); a stateless request is always revision 1

HTTP. Every response carries wrpc-version: the newest revision the server speaks — 2, or 1 for a server that reads no framed messages, which is also what 1.0 answers — named in Access-Control-Expose-Headers so a page on another origin can read it. A client sends a request whose body is a framed message only after a response said 2; until then, and against a 1.0 server always, its bytes travel as JSON. In the other direction a server answers with a framed message only when the request's Accept names application/octet-stream: a 1.0 client, curl and a browser's own fetch name no such thing and read JSON. Accept rather than a request header of wRPC's own because it is CORS-safelisted — a new request header needs a preflight that a 1.0 server, or an application's own cors.headers list, refuses. (A request MAY still send wrpc-version; it is accepted and ignored.) Every request/response pair is self-contained, so a fleet behind one address should agree: an instance that answers 2 promises that every instance reads a frame. One that does not — a 1.0 instance still in the pool, a rollback — refuses a frame before it reads a call out of it, with a wRPC error packet and no wrpc-version: 2; a client that sent a frame and is answered that way speaks revision 1 again and sends the same packets once more as JSON, so the call costs a round trip and is not lost. (An answer with no packet in it — a proxy's error page — changes nothing.)

Worker port. A MessagePort has no handshake, and the two ends are deployed apart — a tab that loaded before a release, a Service Worker that outlives its pages. The page's first packet is a ping naming the revision it speaks, and a 2.x end answers with its own:

{ "type": "ping", "v": 2 }

{ "type": "pong", "v": 2 }

The port speaks the older of the two. A 1.0 end ignores the field and answers a plain pong — revision 1 — and a 1.0 page sends no such ping, so it is sent no framed message. Under a packet codec both are the codec's encoding, like every other packet on the port (and the port is revision 1: a codec owns the wire). The field means nothing on a transport that settled the revision elsewhere.

Changes since 1.0

What a 2.0 peer speaks that a 1.0 peer does not, section by section, and what happens when the two meet. Between a client and a server nothing needs setting: the revision is negotiated, and what is not negotiated is additive. Two rows are not additive, both between the instances of one fleet — where there is no handshake to negotiate on — and each says what to set until every instance is upgraded; the CHANGELOG has the upgrade order.

Since 2.0	Where	Meeting a 1.0 peer	Until every peer is 2.0
Revision 2: framed messages of kind 1 — binary attachments, a packet whose byte values travel as bytes, on WebSocket, a worker port and packet-mode HTTP — never on SSE, which speaks revision 1 to every client, as 1.0 did	Versioning, Binary chunks	Negotiated: a connection with a 1.0 peer speaks revision 1 and carries no frame; its bytes travel as the { "0": … } / { type: 'Buffer', data } JSON of 1.0.	nothing
wrpc-version: 2 on an HTTP response, and the Accept that asks for a framed answer	Versioning	Additive: a 1.0 server answers 1 and is sent no frame; a 1.0 client sends no such Accept and is answered JSON.	nothing behind one address per version; a mixed fleet behind ONE address sets attachments: false on its 2.0 instances, so they answer 1
v on a worker port's first ping/pong	Versioning	Additive: a 1.0 end answers a plain pong, a 1.0 page sends no such ping — revision 1 either way.	nothing
Framed messages of kind 3 and 4 — a packet or a chunk compressed by a Node client — and the enc field of ping/pong that negotiates them	Binary chunks, Compression	Negotiated: a 1.0 server ignores enc on the ping and answers a plain pong, so a 2.0 client never sends one; a 1.0 client never sends one.	nothing
The subprotocol carrier tokens wrpc.h. and wrpc.m. — a browser's declared headers and data on a WebSocket handshake — and a wider deny list of declared names (wrpc.bearer. is 1.0's own)	Connection metadata	A 1.0 server selects wrpc.v1 and reads no token — so a 2.0 browser client that was answered wrpc.v1 dials once more with the wrpc_h/wrpc_meta query 1.0 reads. A 2.0 server that sends no frames answers wrpc.v1 too (and reads the tokens), so the answer is ambiguous: the redial happens there as well, and a client that offered wrpc.v1 alone is never redialled. A 1.0 client never sends a token.	carrier: 'query' on a page that must reach a 1.0 server; carrier: 'protocol' to keep labels out of URLs
The SSE channel secret, presented as x-wrpc-channel: <id>.<secret>	Server-Sent Events	Additive: the ready frame hands out the whole reference in the field 1.0 named channel, and a 1.0 client presents it back as it is. A 2.0 client against a 1.0 server presents the bare id that server gave it.	nothing
epoch and seq on the rooms envelope — loss detection between instances	Rooms	Additive: a 1.0 instance ignores them, a 2.0 instance delivers a 1.0 instance's envelopes untracked.	nothing
The wrpc-bin: rooms and cluster envelope — an event (or a node-to-node question or answer) whose data holds bytes, as an attachments frame	Rooms, Cluster channels	Not additive. A backplane is a broadcast with no handshake: a 1.0 instance cannot parse the envelope and drops the event, without a log line.	attachments: false on the 2.0 instances keeps the envelope JSON, as in 1.0
The wrpc-enc: and wrpc-sealed: rooms and cluster envelopes — compressed, encrypted	Rooms	Opt-in on the publisher; a 1.0 instance drops what it cannot read.	turn rooms.compression, cluster.compression and the encryption options on only once every instance is 2.0 — the two-step rollout the section describes
seq, ch and at on a signed cluster envelope — the sender's counter, channel and clock, under cluster.secret	Cluster channels	Not additive one way. A 1.0 instance verifies a 2.0 envelope as before (the fields are inside the signed bytes) and applies it; a 2.0 instance refuses a 1.0 instance's envelope, which has no counter — the two halves of a mixed cluster stop seeing each other. Accepting one unasked would be a replay hole. Clusters without secret are unaffected.	cluster: { replay: 'accept' } on the 2.0 instances while a 1.0 instance is left
headers and cache on a procedure's http descriptor in introspection	Introspection	Additive: unknown keys of a descriptor are ignored.	nothing
WebRTC: the data-channel framing, the signaling messages, the trust assertions	WebRTC	A new carrier: a 1.0 peer has no WebRTC transport and never meets it.	nothing
The compression rule — enc lists, a sender compresses with the first codec of its own list the peer announced	Compression	Governs only what two 2.0 ends negotiate.	nothing
WebTransport, the broker binding, session encryption with sealed requests and broker sealing — experimental	WebTransport, Broker binding, Session encryption	New carriers and formats, opt-in on both ends; a 1.0 peer never meets them.	nothing

Framing

Transport	JSON packets	Binary chunks
WebSocket	text frames	binary frames
HTTP	request body / response body	not available
SSE	POST body out, data: lines back	not available
Worker port	postMessage(string)	postMessage(Uint8Array)
WebRTC data channel	binary frames, KIND = 0	binary frames, KIND = 1
WebTransport	control stream, KIND = 0	control stream, KIND = 1

A transport that cannot stay open (plain HTTP) carries calls only: events, subscriptions, cancellation and streams need a persistent connection, and asking for one over HTTP is an error rather than a silent no-op.

Besides packet mode (POST {basePath} with a packet body) and the conventional REST mode (ANY {basePath}/:unit/:method, answered with a callback packet), a server MAY serve declaratively mapped routes: a procedure carrying an http descriptor (surfaced through introspection, below) is addressable as method {basePath}{path}, args arrive structured as { params, query, body }, and the response is the plain result (or the wire error object) with the mapped status — external REST semantics. This is additive: a peer that ignores http keeps every existing mode.

Everything on this page describes the JSON framing, which is the protocol. An application MAY replace the framing on both of its own ends with an injected wire codec — that is an opt-in arrangement outside this page's interoperability promise: an independent implementation is only guaranteed to interoperate over JSON. The same holds for REST bodies: both REST modes' bodies (declared routes' plain results and the conventional mode's callback envelopes, requests and errors alike) MAY be re-framed — binary included — by an injected codec.rest section, under the same outside-the-promise terms.

SSE is persistent but text-only, so it carries everything except binary streams — see Server-Sent Events below. A WebRTC data channel is persistent and binary but caps the size of one message, so both packets and chunks travel fragmented under a one-byte header — see WebRTC below. A WebTransport session is persistent and binary but its streams carry bytes with no message boundary, so packets and chunks travel length-prefixed on one stream — see WebTransport below.

Connection metadata

A client may declare connection-phase metadata at connect time: a bag of declared headers and a bag of declared data. This is a transport-level convention, not a packet, and it has one carrier per thing a transport can actually send:

Carrier	Declared headers	Declared data	Who can send it
Real request headers	the headers themselves	x-wrpc-meta (percent-encoded JSON), or one x-wrpc-meta-<key>: <value> per key	HTTP, SSE; a WebSocket client outside a browser
Subprotocol offers	wrpc.h.<base64url>	wrpc.m.<base64url>	a browser WebSocket
Connect-URL query	wrpc_h (percent-encoded JSON)	wrpc_meta (percent-encoded JSON)	any WebSocket client that opts into it; WebTransport

A browser's WebSocket constructor cannot set a request header and fetch refuses to perform the upgrade by hand, so the one handshake header a page controls is Sec-WebSocket-Protocol. A carrier token is the JSON object, UTF-8, base64url without padding — a subprotocol name is an RFC 7230 token, so neither raw JSON nor = is admissible:

Sec-WebSocket-Protocol: wrpc.v2, wrpc.v1, wrpc.h.eyJ4LXRlbmFudCI6ImFjbWUifQ, wrpc.m.eyJ1c2VySWQiOjd9

Carrier tokens — wrpc.h., wrpc.m. and the credential token wrpc.bearer.<token> — are data riding the offer, never a protocol to select. A server MUST NOT echo one: the response would reflect a credential, and the client did not ask for it. Because a client fails a handshake whose offers all went unanswered, a client MUST offer a selectable protocol (a revision — wrpc.v2, wrpc.v1) next to a carrier token, and MUST NOT emit carrier tokens when it offers nothing else. wRPC's negotiators remove the tokens from the offer before an application's protocols/handleProtocols sees it.

wrpc.h. and wrpc.m. are new since 1.0, which read the declared bags from the connect-URL query only and ignores them (wrpc.bearer. is 1.0's own). The revision is how a client finds out: a handshake that offered wrpc.v2, carried a token and was answered wrpc.v1 may have reached a 1.0 server, so a wRPC client dials once more with the query carrier and keeps it for that connection's later reconnects. A client told which carrier to use (carrier: 'protocol' or 'query'), or given its own protocols, is left with that choice.

The x-wrpc-meta-<key> spelling carries string values only and the JSON header wins a key collision. A conformant client MAY emit either; on a carrier without headers the prefixed mode still travels as the JSON bag, because the guarantee is about the bag the server observes, not the wire.

A peer that declares nothing is perfectly normal, and a server that ignores the declarations is conformant. A server that consumes them MUST treat them as untrusted labels, whichever carrier brought them:

• Size-capped on the encoded input, before any decoding. wRPC's default is 2048 bytes: one budget for the two offers together (headers first), the query measured whole, the x-wrpc-meta header on its own. Keep the budget under the smallest request-header limit in the deployment — a WebSocket handshake is an HTTP request, and uWebSockets.js allows 4096 bytes for all of its headers where node allows 16 KB.
• Refused, never fatal. A malformed or oversize declaration leaves the connection unlabelled; it does not close it.
• A carrier is chosen, never merged. For each bag: the real header, else the offer, else the query. An offered token that is refused still silences the query, so garbage cannot downgrade a connection to the other carrier.
• Never able to override an observed request header. Declared names only add, and wRPC drops the names a hostile page could otherwise forge next to a victim's cookie: cookie, host, origin, forwarded, via, x-real-ip, x-client-ip, true-client-ip, cf-connecting-ip, fastly-client-ip, fly-client-ip, remote-user, and everything under sec-, content-, proxy-, x-wrpc- and x-forwarded- — and under the prefixes an identity-aware proxy sets about the user it authenticated: x-auth-request-, x-amzn-oidc-, x-goog-authenticated-user-, x-goog-iap-, x-ms-client-principal. A declared name that is not an RFC 9110 token is dropped, and _ is read as - before the list is checked. Real headers are not filtered — a peer that can send them is not a page, and no deny list binds it.

Keys of both declared bags are normalized to kebab-case (userId → user-id), so one spelling addresses a value whatever carrier brought it and schema.headers has a single casing to validate. Because both spellings reduce to the same key, a collision between them is real rather than two lookalike keys. A conformant server MAY apply this normalization; a caller writing header names by hand must write kebab itself, since HTTP lowercases header names before the server observes them and the word boundary cannot be recovered afterwards.

Packets

Every packet is a JSON object with a type. Unknown types are answered with a callback carrying code 500.

On a persistent connection the packets pair up like this. The sections below take them one at a time:

call — client → server

{ "type": "call", "id": "b1f0…", "method": "chat/send", "args": { "text": "hi" } }

method is unit/name or unit.vN/name (auth.v1/signIn); the version defaults to *. id correlates the answer and is generated by the caller: a string of 1 to 255 characters — the same bound a chunk header puts on a stream's id. A number is taken too, because a 1.0 client with a counting generateId sent one. Any other id — an object, an array, a longer string — is never echoed. Its packet is refused with an id-less 500 (packet.unknown), on every packet type that carries an id.

call, subscribe and event packets may additionally carry an optional meta field — a plain JSON object of caller-declared, unvalidated per-invocation metadata (an idempotency key, a locale, an A/B cohort):

{ "type": "call", "id": "b1f0…", "method": "orders/create", "args": { … },
  "meta": { "idem": "9f3c…" } }

A server that understands it surfaces the (sanitized: plain-object-only, size-capped, frozen) bag to handlers and hooks; one that does not ignores it like any other unknown field. It is a label, never an authorization input, and deliberately outside schema validation. The trace-context fields tp/ts (below) are separate on purpose — they belong to the telemetry propagator, not to the application.

A call packet may also carry an optional timeout field — the caller's per-call deadline in milliseconds (CallOptions.timeout). A server that understands it uses it to shorten the procedure's own time budget (never to widen it — the caller's budget is a courtesy, the procedure's timeout is the server's protection); one that does not ignores it, and only the caller's local timer applies. Additive, absent unless set.

callback — server → client

{ "type": "callback", "id": "b1f0…", "result": { "ok": true } }

{ "type": "callback", "id": "b1f0…", "error": { "message": "Not found", "code": 404 } }

Exactly one of result / error. Error codes reuse HTTP semantics where they fit: 400 invalid input, 403 no session (or a refused origin), 404 unknown method, 408 timeout, 429 too many in-flight calls from one connection, 500 handler failure or invalid output, 503 queue overflow. Over HTTP the same code becomes the response status.

What the message carries depends on the class of the code. 4xx messages are written for the caller — validation text, quota refusals — and travel verbatim. A 5xx message is a server internal: the peer receives the status line ("Internal Server Error") and the exception text stays in the server log, correlated by the same packet id. A server-side error that WANTS its message on the wire opts in with error.expose = true; the errors wRPC itself constructs (timeout, queue overflow, invalid output) are marked so.

The error object MAY carry a third, optional field — details — with structured, JSON-serializable data about the failure. Validation failures use it for their issue list:

{
  "type": "callback",
  "id": "b1f0…",
  "error": {
    "message": "Invalid arguments: text is required",
    "code": 400,
    "details": { "issues": [{ "message": "text is required", "path": ["text"] }] }
  }
}

details obeys exactly the exposure rule message does: it travels on a 4xx (or with expose = true) and is stripped from a 5xx. A peer that does not know the field ignores it, per the additive-fields rule above.

event — both directions

{ "type": "event", "name": "chat/message", "data": { "text": "hi" } }

A plain event is fire-and-forget in both directions: no id, no acknowledgement, no answer. That is what makes it cheap, and it is why a rejected inbound event (unknown handler, missing session, failing handler, invalid input) is recorded in the server log rather than sent back — there is no id to answer on.

The one exception is a transport that cannot carry events at all. An event sent over HTTP is refused with a callback carrying code 400 and an empty id: fire-and-forget on a request/response transport would leave the request unanswered forever, so it is an error rather than a silent no-op.

• server → client: emitted by client.sendEvent(name, data) or by a room broadcast (server.to(room).emit(...)). The client dispatches on the unit part of the name to client.api[unit]; anything that reaches no listener surfaces as an unhandled-event on the client itself.
• client → server: sent by client.sendEvent('unit/name', data) and handled by the unit's on map in the router. Handlers are procedures, so access, input validation and queue all apply.

Asks: an event with an id

{ "type": "event", "name": "chat/confirm", "data": { "q": 1 }, "id": "b1f0…" }

A server → client event MAY carry an id. It then expects an answer: the client MUST reply with an ordinary callback packet carrying the same id — result with what its registered responder returned, or error when it has no responder for the name (501) or the responder threw (the error's code, default 500). No new packet type is involved; the field is additive, and a peer that never sends asks never sees the difference.

Client → server events keep no id — a client asking the server is simply a call. A callback arriving at the server that matches no pending ask is logged and dropped, never answered: answering a callback with a callback could ping-pong.

ping / pong — both directions

{ "type": "ping" }

{ "type": "pong" }

The application-level heartbeat. A browser WebSocket exposes no protocol-level ping, so a connection that died without a close frame — a dropped NAT mapping, a suspended laptop, a proxy that stopped forwarding — looks perfectly open from JavaScript until the first call times out. Both sides answer a ping with a pong immediately; the client additionally measures the round trip and reconnects when the answer does not arrive within its timeout. On a WebRTC link each direction has its own client and its own channel (see WebRTC), so each direction heartbeats on its own channel.

Two optional fields ride the pair, both ignored by a peer that does not know them: enc, the compression offer of a Node WebSocket client, and v, the revision a worker port names on its first ping.

subscribe / data / end / unsubscribe — a stream of values

A subscription is a procedure that answers with many values instead of one. It needs a connection that stays open, so it is refused (400) on HTTP.

{ "type": "subscribe", "id": "b1f0…", "method": "chat/onMessage", "args": {}, "lastEventId": "41" }

{ "type": "data", "id": "b1f0…", "eventId": "42", "data": { "text": "hi" } }

{ "type": "end", "id": "b1f0…" }

end is the terminal packet in every case — normal completion, a generator that threw (error: { message, code, details? } — the same object shape, including the optional details, as a callback error), and refusals such as 404, 403 or 429. {"type":"unsubscribe","id"} ends one early; the server aborts the handler's signal, runs its finally, and answers end.

eventId appears only on values the handler wrapped in tracked(id, data). The client remembers the last one and sends it back as lastEventId when it re-subscribes, which the handler uses to replay what was missed — createEventLog({ size }) is the ring buffer for exactly that. A feed of untracked values simply has no resume point and picks up live.

Server-side backpressure is real: the pump waits for the transport to drain before pulling the next value, so a slow consumer stops the producer instead of filling memory.

cancel — taking a call back

{ "type": "cancel", "id": "b1f0…" }

Cancellation is best-effort by nature — a handler that never looks at ctx.signal keeps running. What is guaranteed is that the caller is rejected immediately with code 499 and that whatever the handler eventually returns is dropped rather than delivered late. Like unsubscribe, it needs a persistent connection.

Batch frames

A JSON array in place of a packet is a batch: several packets in one frame. Each is dispatched and answered on its own, so a failure inside a batch stays attached to its own id.

[{ "type": "call", "id": "a", "method": "math/double", "args": { "n": 1 } },
 { "type": "call", "id": "b", "method": "math/double", "args": { "n": 2 } }]

On a WebSocket the answers come back individually. On HTTP — where a request has exactly one response — they come back as one array in request order, so a caller can zip requests to responses positionally. The server caps a frame at maxBatch packets (128 by default); one frame asking for unbounded work is otherwise a denial of service.

Trace context

call, subscribe and event packets may carry two optional fields that link a client's span to the server's, so one distributed trace spans both sides of the wire:

Field	Carries
tp	W3C traceparent
ts	W3C tracestate; omitted when empty

{ "type": "call", "id": "b1f0", "method": "chat/send", "args": { "text": "hi" },
  "tp": "00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01" }

They are short because they ride on every call packet: "traceparent" would cost eleven more bytes per call than "tp" for nothing.

Both fields are optional in both directions. A peer that does not send one leaves the receiver to start a root span; a peer that does not understand one ignores it, like any other unknown field. The context is per packet, not per frame, so each call in a batch keeps its own parent.

An implementation is not required to parse these — wRPC itself does not. It hands the field to whatever OpenTelemetry propagator the application configured, which is also why a wRPC server only reads tp when it was given the @opentelemetry/api module (or an explicit propagator) rather than a bare tracer.

A server accepts an inbound tp by default, exactly as gRPC and HTTP instrumentations do. Since a hostile peer can forge trace ids, a server that faces untrusted clients can refuse them with telemetry.trustRemoteContext: false and start every trace itself.

stream — both directions

{ "type": "stream", "id": "9a2c…", "name": "video.mp4", "size": 104857600 }

{ "type": "stream", "id": "9a2c…", "status": "end" }

The opening packet announces a stream; status is end or terminate and closes it. The payload travels as binary chunks in between.

Binary chunks

Each binary frame is one chunk of one stream — unless its first byte is 0x00, which no chunk has (a stream id is at least one byte long): such a frame is a framed message, its second byte the kind. This is what revision 2 adds to the core — revision 1 never sent an empty stream id either, but did not reserve the byte, and a 1.0 peer reads such a frame as a chunk of a stream it does not know:

0x00  kind  payload
      1     binary attachments: u32 headerLen (big-endian), JSON `[packet, [[path, byteLength], …]]`, then the buffers back to back
      3     a packet, compressed with the codec negotiated on ping/pong (below); the receiver inflates, then reads it as a text frame
      4     a chunk, compressed likewise; the receiver inflates, then reads it as a binary chunk
      other reserved — a 400 error packet

An attachments frame (kind 1) is a packet — or a batch array — whose byte values travel as bytes: in the JSON every typed array, ArrayBuffer or DataView is replaced by null, and the index names each one by its path from the packet root (object keys and array indexes, the Jupyter buffer_paths shape) with its length; the buffers follow in index order. A receiver MUST refuse a frame whose paths name __proto__, constructor or prototype, walk through a non-container, land on anything but null, exceed 32 levels, or whose lengths do not add up to exactly the bytes that follow — with a 400, never a partial packet — and SHOULD copy the bytes out of the frame rather than hold views into it. The same frame is the body of a packet-mode HTTP request or response — and of a conventional REST answer — under Content-Type: application/octet-stream. It is sent for a packet that holds bytes only where revision 2 was negotiated (Versioning): a WebSocket that selected wrpc.v2, an HTTP exchange whose other side said it reads one, a worker port whose two ends both named revision 2. An end that opted out (attachments: false) or speaks a wire codec negotiates revision 1 in the first place; one that is sent a frame anyway answers it as a malformed packet. SSE, being text-only, speaks revision 1 — bytes as JSON, both ways — and refuses a frame on a channel POST with 415. Kind 2 is reserved.

The negotiation is a ping whose enc is the client's codec ids in its order of preference (a list; one id is a list of one), answered by a pong whose enc is ONE id — the first of the client's list the server holds — or by a plain pong when it holds none. Only the client compresses on this wire, so that id is the codec of every kind 3 or 4 frame that follows. A peer MAY send one only after such a pong; before that, or with a payload that does not inflate under the receiver's cap, the frame is answered with an id-less 400 and the connection goes on. Today only a Node client sends them (a browser's WebSocket compresses both directions under permessage-deflate, and a server's own frames use that extension too).

Each ordinary binary frame is one chunk of one stream:

┌────────┬──────────────────┬──────────────────────────┐
│ 1 byte │  id (idLength)   │        payload           │
│ idLen  │  utf-8 stream id │        bytes             │
└────────┴──────────────────┴──────────────────────────┘

Prefixing the id rather than opening a frame per stream is what lets several streams interleave over one connection without head-of-line blocking. Byte layout and the chunkEncode/chunkDecode helpers are in the wire format reference.

Introspection

Every server answers one procedure it did not declare: system/introspect, injected unless the router already defines it. It is access: 'public' — a client has to be able to ask what it may call before it has a session.

// -> { "type": "call", "id": "1", "method": "system/introspect", "args": ["chat"] }
// <- { "type": "callback", "id": "1", "result": { ... } }
{
  "chat": {
    "send":      { "access": "public", "meta": { "description": "Post a message" } },
    "onMessage": { "access": "session", "kind": "subscription" }
  },
  "auth.v1": { "signIn": { "access": "public" } }
}

args is an array of unit keys to filter by; anything else means no filter (REST-mode calls deliver a plain object here). A unit key is unit for the default version and unit.vN otherwise — the same string client.load() takes.

Each method carries:

Key	Meaning
access	'public', 'session', or whatever the router declared
kind	'subscription', and only then — a client scaffolds a call unless told otherwise
meta	The procedure's meta, when it is not empty. meta.description is the one home for prose — wrpc types turns it into a doc comment, so it does not also live in signature
signature	An optional descriptor, below
http	The declarative REST mapping ({ method, path, status?, headers?, cache? }), when the procedure carries one. headers are the route's static response headers, cache its { maxAge, public, staleWhileRevalidate, etag } policy — both additive, present only when declared.

A unit object may additionally carry two reserved keys — impossible to collide with a method, since both are reserved in the router definition too:

Key	Meaning
on	The unit's inbound (client → server) event handlers: { [event]: { access, signature? } }. What wrpc types turns into the contract's sends key, typing client.sendEvent.
emits	The unit's declared outbound (server → client) events, verbatim from the router's declaration-only emits key: { [event]: { data?, returns? } } in the signature shape language. What wrpc types turns into the contract's events key, typing the unit emitter and client.respond.

This is what load() consumes to build client.api (the two reserved keys are skipped — they are declarations, not methods), and what the wrpc types CLI consumes to generate a contract interface — where a kind: 'subscription' method becomes a SubscriptionContract<Args, Data> member rather than a callable one.

Being an ordinary procedure, it is reachable over every transport: a WebSocket frame, a plain POST {basePath} with the packet as the body (which is how the CLI asks, needing no socket), an SSE channel, a worker port, a WebTransport session, a data channel or a broker request. It also answers in REST mode at {basePath}/system/introspect.

The signature descriptor

signature describes a procedure's arguments and result well enough to generate TypeScript from. It is deliberately a small closed format rather than a schema language: it crosses the network and ends up inside a file someone compiles, so a generator must be able to reject everything it does not recognise.

procedure({
  signature: { args: { room: 'string', 'limit?': 'number' }, returns: [{ id: 'string', text: 'string' }] },
  handler: async (context, { room }) => [...],
})

• args — what the procedure takes. Omitted means undescribed.
• returns — what a call answers with.
• data — what a subscription yields. (returns on a subscription and data on a call are both ignored, with a warning.)

Each of those is a shape, and a shape is one of three things:

Shape	Example	TypeScript
a type name	'string', 'number[]', 'string|null'	string, number[], string | null
a field map	{ id: 'string', 'note?': 'string' }	{ id: string; note?: string }
a one-element array	[{ id: 'string' }]	Array<{ id: string }>

A field name ending in ? is optional. Field maps nest. The type names are exactly string, number, boolean, object, null, unknown, any and never — there is no Date, because there is no Date in JSON — plus [] suffixes and | unions of those. Anything else generates unknown, so a descriptor can never widen what a generator is willing to emit.

Validation is not part of it. A signature says what a procedure looks like; input/output (a function or a Standard Schema) are what actually enforce it, and the two are deliberately separate — a schema library is the user's choice, while this has to survive a JSON round trip.

Sessions

A session is a token plus a state object held in a SessionStore. The token travels in a cookie (HttpOnly; Secure; SameSite=Lax by default), which is read on both HTTP requests and the WebSocket upgrade, so a reconnect restores the session without a round trip.

Because a SameSite=Lax cookie rides along on cross-site navigation, GET/HEAD requests in REST mode only get their cookie-restored session when the request proves same-origin intent through Sec-Fetch-Site. Otherwise they run as public — ambient-authority dispatch would be a CSRF hole.

Rooms

A room is a named set of clients. Broadcasting to one is not a new packet type: it is the ordinary event packet, sent to every member.

client.join('lobby');                     // from a handler: ctx.client.join(...)
server.to('lobby').emit('chat/message', { text: 'hi' });
server.to('lobby').except(ctx.client).emit('chat/message', { text: 'hi' });
server.broadcast('system/announce', { up: true });

emit() returns how many clients received the event on this instance. Rooms are released automatically when a client disconnects. to(...rooms) is a union — a client in several targeted rooms still gets one copy — and to() with no rooms reaches nobody, so a computed room list that came back empty never falls back to every connected client.

Across instances

With a backplane configured (new Server({ backplane }), see @alexify/wrpc/scaling) a room spans every instance sharing it. Each non-local emit is published as an envelope:

{ "v": 1, "instance": "c0ffee…", "epoch": "k3x9…", "seq": 42, "rooms": ["lobby"], "name": "chat/message", "data": { "text": "hi" } }

• instance identifies the publisher, and every instance drops its own envelopes — that is the echo suppression.
• epoch is the publisher's boot marker and seq its per-channel counter: a receiver that sees seq jump within one epoch knows how many envelopes the broker lost between them and reports the gap (backplane.gap, the wrpc.server.backplane.gaps metric). Additive: an envelope without them is delivered untracked. Detection only — the contract stays at-most-once. A receiver compares epochs for equality only: a publisher whose counter table rotated restarts an evicted channel's count under <epoch>.<n>, and the receiver's cursor resets exactly as on a restart.
• rooms is null for a broadcast to everyone.
• An emit targeting exactly one room is published on that room's channel (room:<name>), which only instances holding members subscribe to; anything else goes to the single broadcast channel every instance holds. So an envelope reaches a given instance through exactly one channel and no receiver-side deduplication is needed.

Delivery is at-most-once. A message published while an instance is between subscriptions, or dropped by the broker, is gone: rooms are a fan-out mechanism, not a queue.

With rooms: { compression } on, an envelope past the threshold is published as wrpc-enc:<codec id>:<base64 of the codec's output over the JSON text> — a message that starts with w rather than {. There is nothing to negotiate against on a fan-out, so the marker names the codec: an instance encodes with the FIRST codec of its list and reads any codec on it, and a receiver that holds no such codec drops the envelope (and logs). A deployment turns compression on only once every instance can read it, and changes codec the same way — every instance lists both, then the order is swapped. The cluster channels below do the same under cluster: { compression }, applied after the HMAC signature.

An event whose data holds bytes cannot be a JSON envelope. It is published as wrpc-bin:<base64 of the envelope as a binary attachments frame> — the frame a socket carries such a packet in, over the envelope object instead — and delivered on the other instances as bytes.

With rooms: { encryption } on, every envelope is published sealed:

wrpc-sealed:<kid>:<base64( header ‖ AEAD( frame ) )>
header = u8 version (1) ‖ u8 suite ‖ salt (16) ‖ u64 counter
frame  = u8 flags ‖ [ u8 idLength ‖ codec id ] ‖ body        flags bit 0: compressed, bit 1: body is a binary envelope

kid names the key of the shared keyring (1–32 of A-Z a-z 0-9 . _ -) and travels in the clear; suite is 1 AES-256-GCM, 2 ChaCha20-Poly1305, 255 a cipher both ends were handed. A sender draws salt once per process, derives its own key — HKDF-SHA256(key[kid], salt, "wrpc rooms v1" ‖ 0 ‖ kid ‖ 0 ‖ cipher id) — and counts under it: the nonce is the counter, 64 bits big-endian at the end of the nonce, so no two messages share a (key, nonce) pair, and the salt is drawn again after 2^32 messages. The additional data is "wrpc-sealed v1" ‖ 0 ‖ layer ‖ 0 ‖ kid ‖ 0 ‖ channel, so an envelope moved to another channel, another layer or another key id does not open. Compression happens INSIDE the frame (compress, then seal) and names its codec there; a wrpc-enc: marker never appears inside a sealed envelope. A receiver keeps a sliding window of each sender's counters and drops a repeat — a sender being a (kid, suite, salt), the window in memory only: empty when the receiver starts, and dropped with the least recently heard sender once more than maxSenders (1024) are remembered. The frame carries no time. An envelope that does not open is dropped and logged, never answered. The cluster channels do the same under cluster: { encryption } with the label "wrpc cluster v1", after signing and compression. What stays visible to the backplane: the channel name (and so the room name), the kid, the sender's salt, message sizes and timing.

Cluster channels

The cluster layer (server.cluster — presence, introspection, node-to-node messaging) adds two channels on the same backplane, both held for the life of the process: cluster, which every instance subscribes to, and inst:<instanceId>, one instance's own inbox for request answers and addressed commands. Its envelopes are { v, from, epoch, t, … }, where from is the publisher (self-messages are dropped), epoch is the boot marker that distinguishes a restart from a live node, and t selects the message kind (presence hello/state/delta/bye, request q/answer a, command cmd, node event e). These envelopes are an implementation detail of @alexify/wrpc's own cluster layer, not part of the frozen wire protocol clients speak — they never reach a client connection.

Under cluster: { secret } an envelope ends with sig, the hex HMAC-SHA256 of the envelope's JSON text without it, and carries three fields inside what is signed: seq, the publisher's counter (one per process, across both channels, so a receiver sees gaps), ch, the channel it was published on, and at, the publisher's clock in milliseconds. A receiver verifies the signature, then refuses an envelope whose ch is not the channel it arrived on, whose at is further than its maxSkew (30 s) from its own clock, whose seq it already accepted from that from and epoch (a sliding window of 1024), or whose epoch is a life of from older than the one it follows. An envelope without seq is refused unless the receiver runs replay: 'accept'.

An envelope whose application payload holds bytes — the data of an addressed event (sendTo) or a node event, of a question, or an answer's payload — is published as the binary envelope of Across instances: wrpc-bin:<base64 of the envelope as an attachments frame>, or the same frame inside the sealed one. Its sig is the HMAC of the frame's bytes with sig absent — the receiver deletes sig from the decoded envelope and encodes it again — so a signature made over one form is not valid for the other.

Server-Sent Events

SSE is one-way, so a channel is two halves that find each other by id:

GET  {basePath}/events                                 opens a NEW channel
GET  {basePath}/events  x-wrpc-channel: <id>.<secret>  re-attaches to an existing one
POST {basePath}         x-wrpc-channel: <id>.<secret>  client -> server

The channel id is minted by the server — by the application's generateId, so its format (a uuid, a cuid, a counter) is the application's — and the channel secret drawn by the server: 18 random bytes as base64url. Both are handed out, joined, in the ready frame that opens every stream:

event: ready
data: {"channel":"b1f0….Kx9…"}

ONE string, the channel reference — the secret after the id — which a client treats as opaque and presents back whole in one header; the server splits it at the LAST dot (the secret holds none; the id may). It rides the field 1.0 named channel on purpose: a 1.0 client presents whatever that field held, so it presents the secret without knowing there is one — and a 1.0 server's channel is the bare id, which a 2.x client presents just the same. The secret is the credential. A client cannot propose its own id, and a GET or POST naming an id without the channel's secret answers 409 — the same answer as for an id the server does not hold, so a guessed id learns nothing; 409 is the client's signal to drop its channel state and start a fresh one. (?channel=<id>.<secret> in the query string is accepted for re-attach as well, but the header is preferred — URLs end up in proxy logs.) A value without a dot presents no secret.

On top, the channel is bound to the identity that created it: the session token in the opening GET's cookie (or "anonymous" when it carries none). Every re-attach and every POST must present the same cookie identity — a request that presents the secret without it is refused with 403. Channel creation is also capped (maxChannels, maxChannelsPerAddress): past the caps a new GET answers 503 or 429.

Both halves belong to one server-side client, which is what lets a subscription opened by a POST deliver its values down the stream. A POST answers 202 with no body: every reply, callbacks included, travels on the stream — the same shape the worker port transport has. A POST naming an unknown channel answers 409, like the GET.

Each frame carries the channel's own monotonic id:, and a dropped stream does not destroy the channel. It is held for retention (30 s by default), so a reconnect with Last-Event-ID re-attaches and replays the frames it missed instead of starting over — subscriptions and all. Comment frames (: ping) keep proxies from deciding an idle response is a dead one, and X-Accel-Buffering: no keeps nginx from buffering the stream into oblivion.

Serverless-friendly by construction: no upgrade, no socket beyond the response body, nothing but HTTP in either direction. What it cannot carry is binary — SSE frames are text, so wRPC's binary streams are refused on this transport rather than silently corrupted, and a channel speaks revision 1 to every client: a packet's bytes travel as the JSON 1.0 made of them, both ways, exactly as between two 1.0 ends.

WebRTC since 2.0

Two peers speak wRPC to each other over one RTCPeerConnection carrying two negotiated data channels, one per direction of the protocol's client → server relationship:

channel initiator   (negotiated, id 0 by default)   initiator's client  → responder's host
channel responder   (negotiated, id 1 by default)   responder's client  → initiator's host

The initiator is the peer whose id sorts first (plain string comparison); it makes the offer, and it is the impolite side of perfect negotiation. Both peers create both channels before the first offer, with the same ids and the same label (wrpc) — negotiated channels are not described in the SDP, so the two configurations must agree. The ids are configurable for an application that keeps its own channels on the same connection.

Because each channel is one ordinary client → server wire, every packet on this page travels unchanged: a peer runs a client on the channel it initiates and a dispatcher on the other, and nothing in a packet says which peer is "the server". Signaling (descriptions and ICE candidates) is application-level and reaches the peers through any channel the application chooses — @alexify/wrpc/webrtc ships one over an ordinary wRPC connection — and is not part of the wire either.

What signaling carries is nonetheless fixed, so that peers behind different signalers agree: { type: 'description', description, caps? }, { type: 'candidate', candidate }, { type: 'connect' } (a knock — the non-initiator asking to be dialled) and { type: 'close', reason? }, the goodbye. Its reason is one of goodbye (the application ended the link), refused (the sender would not have the peer: a failed trust assertion or its accept() hook) and gave-up (its redial budget ran out). A receiver MUST end the link on any close, MUST read one without a reason as goodbye, and MUST report one it does not know as unknown rather than refuse it — the set may grow.

Data-channel framing

A data channel message has a size limit (16 KiB is the only size every implementation supports; sctp.maxMessageSize reports what a pair actually negotiated) and wRPC's batch frames and stream chunks are routinely larger. So on a data channel every message is binary (the channel's binaryType is arraybuffer), and a packet or chunk is sent as one or more fragments, each under a one-byte header:

bit 0   KIND   0 = a wrpc packet (UTF-8 JSON — what a WebSocket text frame carries)
               1 = a binary stream chunk (a chunkEncode frame, see the wire-format page)
bit 1   FIN    1 = the last fragment of this message
bit 2   COMPRESSED 1 = the message is compressed with the sender's negotiated codec — only once negotiated
bit 3–7        reserved, MUST be 0

• Fragments of one message are sent back to back on one ordered, reliable channel, so there is no message id and no sequence number: a receiver concatenates fragments until FIN.
• A fragment that is not the last MUST carry at least one payload byte, and a receiver bounds how many fragments one message may arrive in (16 384 by default — its reassembly cap in 1 KiB pieces, never fewer than 1024): the byte cap alone would let a peer cost the receiver a buffered view per byte, or per empty fragment. An empty continuation, or a fragment past the count, is a protocol error like the ones below.
• The KIND and COMPRESSED bits of a continuation MUST equal those of the message it continues; a set reserved bit, a COMPRESSED bit before the two peers' lists shared a codec, a mismatched continuation, a text message that is not valid UTF-8, a reassembly past the receiver's cap (16 MiB by default) or a compressed message that does not inflate under that cap is a protocol error, and the receiver closes the channel — the data-channel analogue of a WebSocket 1002.
• Compression is negotiated through signaling, since the channels have no handshake: a description signal MAY carry caps, a JSON object whose known keys are enc — the peer's codec ids in its order of preference (the rule) — and f, the revision of its two halves as bits: 1 its host reads framed messages, 2 its client does. A client half sends a frame only when the other peer's last description set bit 1 (and it reads them itself); a host half only when bit 2 was set. A peer that announces no f is spoken to at revision 1, both ways. Over a raw channel there is nothing to announce through, so the two applications agree — attachments alike on both ends, as with compression. A peer MAY set the COMPRESSED bit only once the other peer's last description announced a codec it holds; the payload is then the output of the SENDER's chosen codec over the bytes the message would otherwise carry, compressed before fragmentation and inflated after reassembly. Over a channel the application negotiated itself there is nothing to announce through: both ends use the first codec of their list, and whether the bit is in use at all is the application's agreement.
• Fragment size is the negotiated sctp.maxMessageSize capped at 256 KiB, and 16 KiB when nothing is reported. A peer MAY send smaller fragments. A peer whose a=max-message-size is under 1 KiB is refused: the link fails rather than fragment every packet into a thousand pieces.

Everything above the header is exactly the WebSocket wire: the ordering rule "a stream packet precedes the first chunk with its id" holds per channel, ping/pong run per direction, and a subscription resumes with lastEventId across a renegotiated connection the way it does across a reconnected socket.

Trust assertions

Signaling is not part of the wire, but the token a signaling server may attach to it is a format two parties written independently must agree on — a peer verifying, and whoever issues (the wRPC unit or any service that can sign a JWS). An assertion is a JWS in compact serialization (RFC 7515): base64url(header).base64url(payload).base64url(signature).

header   { "alg": "ES256", "typ": "wrpc-rtc+jwt", "kid"?: string }
payload  { "sub": string,      the peer id, as the signaling layer names it
           "fp":  string,      the DTLS certificate fingerprint: "sha-256 AB:CD:…"
           "iat": number,      seconds since the epoch
           "exp": number,
           "iss"?: string,
           …any other claims }

• alg MUST be ES256 (ECDSA over P-256 with SHA-256; the signature is the raw 64-byte r || s, as JWS specifies). A verifier MUST refuse any other alg, none included, and any typ but wrpc-rtc+jwt.
• kid selects the public key when the issuer publishes several; a verifier configured with a single unlabelled key accepts any kid.
• Binding. An assertion travels inside a description signal, as its assertion field, and is valid for that description only: sub MUST be the peer id the signaling layer reports as the sender, and fp MUST be the fingerprint the description's SDP declares — everya=fingerprint: line of it, session- or media-level, of any algorithm, MUST equal fp (algorithm, a space, colon-separated hex — compared after normalization: algorithm in lower case, hex in upper case). DTLS binds to the media-level line (RFC 8122 §5), so a description naming a second certificate anywhere is one the relay may have edited, and a verifier MUST refuse it. The DTLS handshake then proves the sender holds that certificate. A verifier MUST apply the description only after the assertion verified.
• exp MUST be in the future by the issuer's clock (a verifier SHOULD allow a small skew and MAY learn the issuer's clock from tokens issued to itself). iss, when the deployment sets one, MUST match.
• A token is at most 4 KiB. Anything else is a refusal; the reason is the verifier's business, not the wire's — the link is simply closed.

WebTransport since 2.0

A client speaks wRPC to a server over a WebTransport session (HTTP/3) the way it does over a WebSocket: the packets are the same and travel in the same order. This section is experimental: it describes revision 1 of the carrier, may change in a minor release, and sits outside this page's interoperability promise until it stabilizes — the way an injected wire codec does.

The client establishes the session (an extended CONNECT with :protocol webtransport, per the WebTransport specification) and opens one bidirectional stream, the control stream; the server treats the first bidirectional stream the client opens as such, and a session that opens none within the server's accept window is closed with code 408. Every packet and every stream chunk, in both directions, travels on the control stream. The session's datagrams carry unreliable events, and — once both ends have announced it — a binary stream's chunks travel on a stream of their own; both are described below.

Connection metadata travels as on a browser WebSocket: the CONNECT request's headers are the observed bag, and the wrpc_h / wrpc_meta query parameters of the request path carry what the client declares. A CONNECT carries no cookies (its credentials mode is omit), so a session token is presented as a declared authorization header or a payload field — what the bearer and payload session transports read.

Stream framing

A QUIC stream is a byte stream with no message boundaries, so every message on the control stream is prefixed with a five-byte header:

bytes 0–3  LENGTH  payload byte length, unsigned, big-endian
byte  4    KIND    0 = a wrpc packet (UTF-8 JSON — what a WebSocket text frame carries)
                   1 = a binary stream chunk (a chunkEncode frame, see the wire-format page)
                   2 = a capabilities message (UTF-8 JSON) — see below
                   3 = a packet, compressed with the sender's negotiated codec — only once `enc` was negotiated
                   4 = a chunk, compressed — likewise
                   5–255 reserved

• A message is one contiguous run of bytes; there is no fragmentation and no FIN bit, because the stream is ordered and reliable. A receiver buffers until LENGTH bytes have arrived, however the transport split them.
• A reserved KIND, a LENGTH past the receiver's cap (16 MiB by default) or a packet that is not valid UTF-8 is a protocol error: the receiver closes the session with code 1002 — the WebTransport analogue of a WebSocket 1002.
• Close codes are the WebSocket close codes carried in the session's closeCode: a server closes with 1001 on shutdown and 1002 on a protocol error; an application close is 1000 or 0. Ending the control stream (a FIN) ends the connection, answered by a session close with 1000.

Everything above the header is exactly the WebSocket wire: the ordering rule "a stream packet precedes the first chunk with its id" holds, ping/pong run per direction, and a subscription resumes with lastEventId across a new session the way it does across a reconnected socket.

Capabilities and per-stream transport

Each end's first message on the control stream MAY be a capabilities message (KIND 2): a JSON object whose known keys are streams ({"streams":true}), enc (the end's codec ids in its order of preference, ["deflate-raw"] for the platform default) and f (1: this end reads framed messages — revision 2). Unknown keys are ignored; an end that sends none, or {}, has announced nothing, and its peer treats it as a revision-1 peer.

The revision is settled here, as on a worker port: an end sends a framed message only once the peer's capabilities carried f, and only when it reads them itself — so a session speaks revision 2 when both ends do, and an end with attachments: false or a packet codec announces no f and is sent none. Until the peer's capabilities arrive, a packet holding bytes leaves as the JSON 1.0 made of it, which every end reads.

Once the two enc lists share a codec, either end MAY send a packet as KIND 3 or a chunk as KIND 4: the payload is the output of the SENDER's codec (the rule — no frame names it) over the bytes KIND 0 or 1 would have carried, and the receiver inflates it before reading — a KIND 3 payload is UTF-8 JSON only after inflation. The choice is per message and the sender's (a message under its threshold, or one the codec did not shrink, goes as KIND 0 or 1 at any time). A compressed kind before the lists shared a codec, a payload that does not inflate, or one that inflates past the receiver's cap is a protocol error (1002). Datagrams and chunks on their own streams are never compressed.

When both ends announced streams, a sender MAY carry a binary stream's chunks on a unidirectional WebTransport stream of their own instead of the control stream:

• The stream packet that opens the wRPC stream ({ type: 'stream', id, name, size }) stays on the control stream — its order against the call that names the id is what the control stream guarantees.
• The unidirectional stream opens with the chunk header — one byte of id length, then the id (exactly the prefix of a chunkEncode frame) — once, and then carries the payload bytes of every chunk of that id, in order, with no per-chunk header. Its FIN is the stream's end; a RESET (an aborted stream) is its terminate. The sender MUST NOT also send the end or terminate packet for such a stream on the control stream.
• A QUIC stream is ordered only against itself. A receiver MUST therefore hold chunks that arrive before their stream packet has been read from the control stream, and MUST deliver the synthesized end or terminate only after the unidirectional stream has ended — never on the strength of the control stream alone.
• The choice is per stream and the sender's: a stream whose chunks were sent on the control stream ends with an end packet there, exactly as in revision 1, and a receiver accepts both forms at any time.

Datagrams

A session's datagrams carry unreliable events: a packet a sender chose to deliver at most once, unordered, for state a later packet supersedes (a cursor, a position). A datagram is one whole message under a one-byte header — the KIND byte alone, since a datagram announces its own length:

byte 0   KIND   0 = a wrpc packet (UTF-8 JSON); other values reserved
bytes 1…        the packet

• Only event packets without an id MAY be sent as datagrams: a call, a callback, a stream packet, a chunk or an ask expects an order or an answer that a datagram cannot promise. A receiver handles a datagram's packet exactly as it would the same packet from the control stream.
• A sender MUST NOT split a packet over datagrams: a packet larger than the session's maxDatagramSize goes on the control stream instead. The choice is the sender's alone and invisible to the receiver; a peer with no datagrams sends every event on the control stream.
• A receiver ignores a datagram it cannot read (a reserved KIND, invalid UTF-8, an empty one) — a lost datagram is the norm, and an unreadable one is not worth a hangup.
• A sender SHOULD drop, not queue, a datagram its session is not ready for: a datagram that waits is a stale one. A dropped datagram is reported to the application as sent — it is not retried on the control stream, where it would wait behind the same congestion.

Broker binding since 2.0

A client speaks wRPC to a server through a message broker — RabbitMQ, NATS, Redis — instead of over a socket: the packets are unmodified wRPC packets, carried as broker messages with a few headers around them. This section is experimental: it describes revision 1 of the carrier, may change in a minor release, and sits outside this page's interoperability promise until it stabilizes, like the WebTransport carrier above.

The broker supplies addressable inboxes with at-most-once delivery, per-sender ordering and optional competing groups (the direct capability, Message brokers). A message carries a body, string headers, a correlationId and a replyTo address. The binding's own headers all start with wrpc-; a peer's connection headers (authorization, x-wrpc-meta, …) ride next to them, lower-cased, and any peer-supplied wrpc-* name is dropped by the receiver.

Header	Values
wrpc-kind	request, response, hello, welcome, packet, chunk, bye
wrpc-seq	a session frame's number, from 1, per direction
wrpc-inbox	on welcome: the address the session's frames go to
wrpc-reason	on bye: why, for logs
wrpc-enc	on a request, hello or welcome: the sender's codec ids, joined by commas, in its order of preference (zstd,deflate-raw); on a response, packet or chunk: the ONE codec its body is the output of — the sender's first that the other end listed. A welcome carries it only when the lists share a codec. A marked frame a receiver cannot inflate ends the session.
wrpc-version	on a hello: 2 when the client reads framed messages; on a welcome: 2 when the service does. A session sends a frame only to a side that said 2 (revision); a stateless request is revision 1 both ways — its body is read as text and its answer is JSON

Every server instance consumes one service address — wrpc.<service> by default — as members of one competing group, so each message addressed to the service reaches exactly one instance.

Stateless requests

client --request {correlationId, replyTo = client inbox}--> service address
client <--response {correlationId}------------------------- the instance that took it

The request body is exactly what a packet-mode HTTP POST carries — one packet or a batch frame — and the response body exactly what its answer carries. The request's peer headers are that POST's headers: session tokens, declared metadata and trace context are read from them the same way. Any instance answers any request; nothing binds a client to one. A request that cannot carry its answer on a later frame — a subscribe, a cancel, an event — is refused as it is on HTTP, with code 400.

Sessions

client --hello {correlationId = session id, replyTo = client inbox}--> service address
client <--welcome {wrpc-inbox = instance inbox}------------------------ the instance that took it
client --packet | chunk {wrpc-seq, correlationId}--> instance inbox
client <--packet | chunk {wrpc-seq, correlationId}-- the instance
either --bye {correlationId}--> the other

• The session id is chosen by the client and names the session in every later message, in both directions (correlationId). The hello's peer headers are the session's connection metadata, as a WebSocket upgrade's headers are.
• A packet body is one packet or a batch frame, UTF-8; a chunk body is a binary stream chunk (a chunkEncode frame, see the wire format). Everything a WebSocket carries — calls, events, subscriptions, cancellation, streams — travels this way.
• Frames are numbered from 1 in each direction. A receiver that sees a number other than the next one treats the session as lost: the carrier is at-most-once, and a missing frame is a missing packet nobody would ever answer. The server ends the session; the client closes and reconnects.
• A frame for a session the instance does not hold (it restarted, or the session idled out) is answered with a bye to its replyTo, and the client reconnects.
• The server ends a session nothing arrived on for its idle timeout. The client's app-level heartbeat (ping/pong) is what keeps a live session inside it.
• A server that refuses sessions answers hello with bye.

A client that reconnects says hello again and may be welcomed by another instance; its subscriptions resume with lastEventId as on any reconnect.

Sealed frames and events

A broker keeps what it carries — for a topic's whole retention, for whoever can read it or its backups. With encryption: { keys } on the binding (the service and its clients share the keyring, as instances share rooms.encryption) every message is sealed with the backplane envelope's frame:

headers   wrpc-sealed: <kid>     + wrpc-kind, wrpc-seq on an RPC frame
body      sealed( u32 headerLength ‖ JSON(headers) ‖ body )

The HEADERS move inside with the body — a client's authorization and x-wrpc-meta, the trace context, wrpc-enc, wrpc-inbox, wrpc-reason — so a bearer token no longer rests in the broker. An RPC frame is bound to address ‖ 0 ‖ kind ‖ 0 ‖ correlation id ‖ 0 ‖ seq (label "wrpc broker-rpc v1"): the readable wrpc-kind and wrpc-seq cannot be rewritten, a frame does not open in another conversation, and a per-sender window drops a replay. A published event is bound to its topic (label "wrpc broker-log v1"), carried as base64 text — a log is only promised to keep a string — and has no replay window: a log is read again and a delivery redelivered, by design. The partition key stays readable; the broker routes by it. A message that does not open is dropped (RPC), skipped (a feed) or dead-lettered with 400 (a consumer), logged, and never answered. A session opened by a sealed hello — welcomed by a sealed welcome — takes sealed frames only from then on, whatever acceptPlaintext says: a plaintext packet, chunk or bye naming it is a downgrade, dropped on either end without consuming a sequence number.

A sealed request and a sealed hello — the frames that reach the competing group — carry wrpc-t, the sender's clock in milliseconds since the epoch, inside the seal. A receiver MUST refuse one whose clock is further than its skew allowance (five minutes by default) from its own: the envelope's per-sender replay window lives in one process, and a captured frame replayed to another instance would otherwise be new to it. A receiver MAY consult a memory shared by the instances (keyed by the envelope's own header — key id, sender salt, counter) to refuse the replay within that allowance too.

Compression since 2.0

Every transport carries plain bytes unless the application turns compression on; nothing negotiates it by default. Where it exists it is the carrier's own mechanism, never a field of a wRPC packet:

Transport	Mechanism	Negotiated by
WebSocket	RFC 7692 permessage-deflate (perMessageDeflate on the engine)	the upgrade handshake
HTTP, packet mode and REST	Content-Encoding on the response — gzip by default, br, zstd or an application's coding by http.compression.encodings	the request's Accept-Encoding against the server's list, in the server's order; the response carries Vary: Accept-Encoding
Server-Sent Events	Content-Encoding on the stream — one encoder for the response (for gzip, one member), flushed after every event (sse.compression)	the opening GET's Accept-Encoding, per response
WebTransport	per message, KIND 3/4 on the control stream (compression on both ends)	the enc list of the capabilities message
WebSocket, client → server from Node	per message, framed binary (0x00 03 / 0x00 04) above the extension (compression on both ends)	{ type: 'ping', enc: [ids] } from the client, answered by { type: 'pong', enc: id } — the first of them the server holds
WebRTC	per message, the COMPRESSED bit of the data-channel header (compression on both peers)	the caps.enc list of the description signal; a raw channel by the application's agreement
The broker binding	per frame, wrpc-enc on the frame (compression on both ends)	the wrpc-enc lists of hello/welcome for a session; a stateless request lists what it accepts and only the answer is compressed
The rooms backplane, the cluster channels	the whole envelope, base64 under a wrpc-enc:<id>: marker (rooms.compression, cluster.compression)	none — the marker names the codec; every instance must hold it, in a two-step rollout

Choosing the codec. A codec is named by an id — deflate-raw (raw DEFLATE, RFC 1951, no zlib or gzip wrapper), brotli (RFC 7932), zstd (RFC 8878), deflate-raw+dict:<hash> for a preset dictionary, or an application's own; an id holds no comma and no whitespace. Wherever two ends negotiate, each announces the ids it can decode, in its order of preference, and a sender compresses with the first codec of its own list the other end announced. The receiver holds both lists, so it knows which codec that is: no frame names it (the broker binding's frames do, its header being there anyway), the two directions choose independently — a server may answer in zstd a client that sends deflate-raw — and there is no tie to break. Lists that share nothing, or an end that announced nothing, leave the wire plain. A receiver looks at no more than the first 16 ids of a peer's list and ignores entries that are not strings.

An HTTP response is encoded only when its body is at or over the configured threshold and nothing upstream already set a Content-Encoding; a 204 and a 304 never are. A REST route's weak ETag is computed over the plain body, so it is the same validator whichever encoding the peer asked for. On an event stream the encoding is a property of one response: a re-attach negotiates it again, and a replay goes out in whatever the new response negotiated.

Session encryption since 2.0

This section is experimental — this one, Sealed requests under it and Broker sealing: they describe revision 1 of each format, may change in a minor release, and sit outside this page's interoperability promise until they stabilize, the way an injected wire codec does. The subpath is marked @experimental whole (Stability).

Opt-in (@alexify/wrpc/encryption), on the persistent transports, and never in place of TLS. A client announces it with wrpc_e=1 in the connect URL — the server may be the first to send, so it has to know the mode before any frame. The flag is not a secret: a client configured to encrypt never accepts plaintext, so stripping it buys a refusal.

Every frame of such a connection is a binary framed message:

handshake, first   00 05 ‖ u8 version (1) ‖ u8 nameLength ‖ protocol name
                         ‖ u8 kidLength ‖ kid ‖ noise message
handshake, later   00 05 ‖ noise message
sealed             00 06 ‖ AEAD( u8 inner ‖ payload )        inner: 0 text, 1 bytes

The handshake is the Noise Protocol Framework (revision 34) under its canonical names — Noise_<pattern>_25519_<AESGCM|ChaChaPoly>_SHA256, patterns NN, NK (the default: the client pinned the server's static key), XX (mutual) and NNpsk0. The first message names the protocol and the key id of the server static it was written for; the server holds it or closes the connection. Nothing is negotiated back — a reply of "try this instead" would be a downgrade channel. The prologue is "wrpc.v1" ‖ 0 ‖ transport kind ‖ 0 ‖ hello header, so a header rewritten in flight, or a handshake replayed onto another kind of connection, fails the handshake. (That wrpc.v1 is the label of this framing, fixed with its own version byte — not the revision the connection negotiated.)

After it, each direction is a Noise CipherState: the nonce is the message counter (four zero bytes, then 64 bits — big-endian for AESGCM, little-endian for ChaChaPoly), the additional data is empty, and both ends rekey (REKEY, Noise §11.3) every 2^20 messages. The interval is a per-deployment constant (rekeyAfter, 2^20 by default), not negotiated — no handshake message carries it — and two ends that disagree fail at the first rekey, as a frame that does not open. Text packets, stream chunks and the other framed kinds all travel as the payload of a sealed frame, so compression happens inside it. A frame that does not open — altered, replayed, out of order — closes the connection: 1002 for anything that did not verify or parse, 1008 for a refusal of policy (an unlisted protocol, an unknown kid, authorize, the handshake timeout, plaintext under required). The close reason is always encryption; which check it was is in the server's log only.

A server's static keys are derived from one configured secret per kid — X25519(HKDF-SHA256(secret, "wrpc noise static v1")), and a second pair under "wrpc hpke static v1" for the per-request binding — and published as the bundle <kid>:<noise key>:<hpke key> (base64url), which is what a client pins.

On WebTransport the same frames ride the control stream as KIND_BINARY messages, after the capabilities exchange, with wt as the transport kind in the prologue. A sealed session announces no streams capability and uses no datagrams — binary streams fall back to chunks on the control stream and an unreliable event is sent reliably — because a stream or a datagram of its own would be a way around the channel; the carrier's own per-message compression is left off as well, since what it would compress is ciphertext.

Sealed requests

A request is not a connection, so the HTTP transport seals each one on its own with HPKE (base mode, DHKEM(X25519, HKDF-SHA256), HKDF-SHA256, and AES-256-GCM or ChaCha20Poly1305) to the server's hpke static key — the construction of Oblivious HTTP (RFC 9458 §4), without its relay:

POST <the transport's endpoint>        Content-Type: application/wrpc-sealed
request    u8 version (1) ‖ u8 aead id ‖ u8 kidLength ‖ kid ‖ enc (32) ‖ ct
plaintext  u32 headerLength ‖ JSON { m, u, h, t } ‖ body
response   nonce (32) ‖ AEAD( u32 headerLength ‖ JSON { s, h } ‖ body )

The real request is inside — method m, path and query u, headers h, the sender's clock t in ms — so an observer sees one endpoint being POSTed to, and a GET REST route travels as that POST too. The inner headers are laid over the outer request's, with three exceptions a server MUST keep: what the connection or a proxy in front of it says about the sender (host, origin, forwarded, via, x-forwarded-*, x-real-ip, the CDNs' client-ip spellings, sec-*) and what an identity-aware proxy says about the user (remote-user and the x-auth-request-*, x-amzn-oidc-*, x-goog-authenticated-user-*, x-goog-iap-*, x-ms-client-principal* names — the same as on the handshake) is never taken from inside — its absence is a fact too, so an inner one is dropped whether or not the outer request carried it; the framing names (content-length, transfer-encoding, connection) are the outer request's; and an outer cookie wins over an inner one. headerLength above 16 KiB is refused. HPKE info is "wrpc http v1" ‖ 0 ‖ the request prefix (version, AEAD id, kid). The answer is sealed under a key both ends export from the same context: secret = Export("wrpc http response", Nk), prk = Extract(enc ‖ nonce, secret), then Expand(prk, "key") and Expand(prk, "nonce"). The outer status is always 200; the real status s and headers h are inside, except Set-Cookie, which stays on the outer response because script cannot set an HttpOnly cookie.

HPKE has no replay protection, and the party this exists for is exactly the one who could replay: a request whose t is further than maxSkew (5 min) from the server's clock, or whose enc was already accepted, is refused 409. Anything that does not parse or open is refused 400. A refusal is never sealed and never detailed — and a client that encrypts treats ANY response that is not application/wrpc-sealed as an error, whatever its status.

Server-Sent Events ride the same binding: the stream request and every channel POST are sealed requests — so the channel id and Last-Event-ID travel inside — and the stream comes back as

Content-Type: text/event-stream; wrpc-sealed=1; n=<base64url(nonce32)>
data: <base64( AEAD( one chunk of the real stream ) )>

one opaque data: event per chunk of the real stream, control frames (ready, gap, the heartbeat comment) included. The key is Expand(Extract(enc ‖ n, Export("wrpc sse stream", Nk)), "key", Nk): the secret only the two ends of the request that opened THIS stream can export, salted with the request's enc and 32 bytes n the SERVER draws per stream and carries in the Content-Type parameter (the one response header a cross-origin page can always read; a stream without a 32-byte n MUST be refused). The frame nonce is the counter from zero. So a re-attached stream has a new key and replays under it, and so does the same open request replayed onto another instance, whose replay memory is its own — without n the two streams would count from zero under one key. A frame that is dropped, reordered or altered errors the stream, which the client sees as a broken connection and re-attaches from. The type stays text/event-stream so that intermediaries treat it as one; an HTTP content coding is never applied to a sealed stream or a sealed answer.

GET <basePath>/encryption-key answers { "key": "<bundle>" } — the one plaintext answer under required (discovery: false removes it). It is trust on first use.

What stays outside the sealed channel is everything the upgrade carried: the URL, real headers, the wrpc.bearer. / wrpc.h. / wrpc.m. subprotocol tokens, cookies. Under session encryption a credential belongs in authenticate or a per-call meta, not in the handshake request.

Reconnect

The client reconnects on its own with truncated exponential backoff and full jitter:

delay = random(0, min(maxDelay, minDelay * factor ** attempt))

Jittering the whole window rather than adding a small offset is what breaks up the thundering herd — after a server restart, a thousand clients that disconnected in the same millisecond would otherwise all come back in the same millisecond.

On a successful reconnect the client reloads every unit it had loaded (the new connection is a new server-side client, so its introspected method list has to be rebuilt) and then emits reconnect. The api unit objects themselves are reused, so event listeners registered on them survive the outage.

Re-authentication is application-level and carries nothing new on the wire: a client configured with an authenticate hook issues ordinary call packets (its credential leg) after the socket opens and before it re-sends its subscribe packets and re-introspects — packets on one connection are ordered, so the server observes the credential first. A server needs no support for this beyond answering the calls.