Streams

Use this skill when reaching for any I/O in Pulp: reading files, sending bytes over a socket, polling an HTTP endpoint, piping between processes, or wrapping a new transport. The pulp::runtime::Stream / AsyncStream hierarchy is the sanctioned path; do not add new Socket::send / http_get callers unless you have a specific reason.

Decision tree

You need to...	Use
Read or write a local file	FileStream (sync) — or wrap in AsyncStream for large files
Keep bytes in memory for tests / round-trips	MemoryStream
Talk to another process via pipe	PipeStream around NamedPipe
Open a TCP connection	TcpStream, wrapped in AsyncStream so connect() doesn't block the caller
Fetch an HTTP/S body	HttpStream::get(url) / post(url, body)
Fetch with custom headers or consume an incremental HTTP/S response	http_request(HttpRequest{...})
Send structured WebSocket frames	WebSocketChannel::connect() / ::accept() over a TcpStream
Send/receive OSC messages	OscChannel::open(host, remote_port, local_port)
RPC-style request/response over any transport	JsonRpcPeer wrapping any MessageChannel
In-process message bridge (tests, inspector)	MemoryMessageChannel::make_pair()

NamedPipe / PipeStream gotchas

On POSIX, NamedPipe presents one public pipe name but uses paired FIFOs internally so bidirectional users cannot read back their own writes. Keep that invariant when changing NamedPipe: a single O_RDWR FIFO looks convenient, but InterprocessConnection read threads can consume outbound frames from the same process and make child-process IPC flaky.

The paired FIFOs must stay directional after connection: server reads the public FIFO and writes the .reply FIFO; the client does the opposite. Do not keep a local writer open just to simplify setup, because EOF/HUP then stops representing peer death. On macOS, protect FIFO write descriptors with F_SETNOSIGPIPE; otherwise killing a connected child can terminate the parent with SIGPIPE.

NamedPipe::read() also has to unblock promptly when close() is called from another thread. Use bounded polling or an equivalent wakeup path; do not leave a background IPC reader stuck in a blocking FIFO read while InterprocessConnection::disconnect() is trying to join it.

InterprocessConnection treats transport EOF as the normal disconnect signal. Directional POSIX FIFOs let a bare NamedPipe::read() report EOF when the peer closes or exits without sending protocol data; keep raw peer-close and abrupt child-exit regression tests in place.

AsyncStream — the patterns that actually work

1. Dispatch callbacks onto your own loop

AsyncStream never links pulp::events directly (that would be a library cycle). Pass an executor closure:

AsyncStream::Options opts;
opts.executor = [loop](std::function<void()> fn) { loop->dispatch(std::move(fn)); };

Without an executor, callbacks run on the AsyncStream's worker thread — fine for tests, usually wrong for UI state.

2. Backpressure: check the return of write_async

write_async returns false when the pending byte count would exceed options.write_high_water (default 1 MiB). When false, wait for on_drain before retrying. Ignoring the bool silently drops the write.

3. Cancelling drains queued writes

cancel() and stop() both complete any queued write callbacks with StreamError::Closed. Do not write code that assumes a cancel "silently forgets" in-flight writes — the callback will fire.

4. Restarting after cancel resets the token

AsyncStream::start() clears a previously cancelled token before launching workers. A stream can therefore be cancelled, stopped, and started again for tests or reusable transports. Do not preserve a stale cancelled token across restart.

5. Null writes are invalid unless size is zero

write_async(nullptr, 0, cb) is the normal zero-byte no-op and completes successfully. write_async(nullptr, nonzero, cb) queues the callback with StreamError::Invalid instead of dereferencing the pointer. Keep that distinction when adding transport wrappers.

6. Writes and auto-reads run on separate threads

When options.auto_read = true the reader and writer run on separate threads so a blocking read() on a TcpStream cannot starve queued writes. Do not re-introduce a single worker loop without understanding this — request/response flows break otherwise.

7. Callback state must outlive the stream and executor

When callbacks capture mutexes, condition variables, or other local state, declare that state before the AsyncStream and before any executor loop that may run queued callbacks. stop() / destruction can dispatch on_close; do not let callback captures die before the stream and loop have drained.

8. Incremental HTTP callbacks run inline

HttpRequest::on_chunk runs synchronously on the thread calling http_request(). Return false to abort the transfer. A request with a chunk callback leaves HttpResponse::body empty, so either consume bytes in the callback or omit the callback for buffered behavior. Chunk boundaries are transport boundaries, not message or SSE-event boundaries; retain incomplete protocol input between callbacks. Keep credentials in HttpRequest::headers; the runtime reports generic transport and callback errors without reflecting request header values.

Extending with a new transport

To add WebSocket, S3, or any other transport:

1. Implement Stream::read / write / close / is_open.
2. Return StreamResult::fail(StreamError::WouldBlock) from read() when no data is available; AsyncStream handles backoff.
3. Keep read() non-blocking if at all possible — if it must block, document it so callers know to always wrap in AsyncStream with auto_read = true.

Message channels

MessageChannel is the structured-message layer: one send() = one delivered message. Use it when the peer protocol doesn't tolerate partial reads (WebSocket, OSC, JSON-RPC, etc.). Callback dispatch follows the same executor contract as AsyncStream.

Patterns that are easy to get wrong:

1. WebSocket handshake failure returns nullptr. WebSocketChannel::connect and ::accept both return an empty unique_ptr on a bad handshake — do not dereference blindly.
2. OSC is UDP; packets can be dropped or reordered. For anything that needs reliability, layer JsonRpcPeer over WebSocketChannel, not OscChannel.
3. JsonRpcPeer is symmetric. Either side can register methods, send requests, and fire notifications; a client/server split is only a convention.
4. JSON-RPC params are JSON strings, not choc values. This keeps the public surface free of CHOC types. Format with choc::json::toString on the way in and choc::json::parse on the way out if you need structured access.
5. Interrupt a WebSocket reader before releasing its socket. WebSocketChannel::close() uses TcpStream::shutdown() to wake blocking reads and mark the stream closed. Destruction then joins the reader before releasing the socket handle with close(). Preserve that order; closing the handle while the reader is inside receive() races with descriptor invalidation.
6. Do not destroy a WebSocket channel from an inline callback. With no executor, callbacks run on the reader thread. They may call close(), but must defer destruction until after the callback returns; otherwise the destructor would attempt to join its own reader. Use an executor when the callback needs to own channel lifetime.

References

• Docs: docs/reference/streams.md (full API + backpressure flow + MessageChannel)
• Headers: core/runtime/include/pulp/runtime/{stream,async_stream,network_stream,message_channel,websocket_channel,memory_message_channel,json_rpc}.hpp, core/osc/include/pulp/osc/osc_channel.hpp
• Example: examples/stream-demo/main.cpp
• Tests: test/test_{stream,async_stream,network_stream,websocket_channel,osc_channel,json_rpc}.cpp — copy these patterns for new transport tests
• Feature plan: planning/next-features-plan.md stream feature background