alknet/tasks/call/protocol/dispatch-streaming-branch.md

id, name, status, depends_on, scope, risk, impact, level

id	name	status	depends_on	scope	risk	impact	level
call/protocol/dispatch-streaming-branch	Wire Dispatcher::handle_stream streaming branch (Subscription → invoke_streaming → write each → call.completed)	pending	call/registry/invoke-streaming	narrow	medium	component	implementation

Description

Wire the server-side streaming dispatch branch in Dispatcher::handle_stream / dispatch_requested. When a call.requested arrives for a Subscription op, the dispatcher must call OperationRegistry::invoke_streaming() and pump the resulting ResponseStream to the wire: each Ok(value) → call.responded frame, Err → call.error frame (terminal), natural stream end → call.completed frame. This is the server-side path that makes Subscription operations work end-to-end — without it, a StreamingHandler-registered op had no server-side dispatch path.

This task depends on call/registry/invoke-streaming (which provides invoke_streaming()). It adds the op_type branch to the dispatch path and the stream-to-wire pump.

The branch

dispatch_requested currently unconditionally calls registry.invoke() and returns one ResponseEnvelope. It needs to know the op_type to branch. Two options:

1. Look up the registration to get op_type before dispatching. The build_root_context already looks up the registration; expose op_type from it. Then branch: Subscription → streaming path, Query/Mutation → existing invoke() path.
2. Return an enum from dispatch_requested (DispatchResult::Once(ResponseEnvelope) | DispatchResult::Stream(ResponseStream)) and let handle_stream match on it for the wire-writing loop.

Pick the cleaner option. Option 1 keeps dispatch_requested returning a ResponseEnvelope for the Once path but needs a separate streaming entry point (e.g., dispatch_requested_streaming returning ResponseStream). Option 2 unifies the dispatch entry but changes the return type. The spec frames it as "branches on op_type" in handle_stream, suggesting the branch lives in the dispatch layer. Document the choice.

Streaming dispatch path

For a Subscription op:

// In dispatch_requested (or a new dispatch_requested_streaming):
let context = self.build_root_context(...);
// deadline: None for subscriptions (unbounded — ADR-049 §6, call-protocol Timeouts)
// The build_root_context sets a 30s deadline; for the streaming path, set
// deadline to None AFTER construction (or pass a flag). The spec says
// "deadline: None for subscriptions (unbounded)".
let stream = self.registry.invoke_streaming(&operation_name, input, context);
stream  // ResponseStream — pumped by handle_stream

handle_stream streaming pump

In handle_stream, after dispatching, if the result is a stream:

// Read the ResponseStream, write each envelope as an EventEnvelope frame
let mut stream = tokio_stream_into_response_stream(...);  // or use StreamExt
while let Some(envelope) = stream.next().await {
    let event: EventEnvelope = envelope.into();
    if let Err(err) = writer.write_frame(&event).await {
        warn!(error = %err, "failed to write streaming frame; closing stream");
        break;
    }
    // If the envelope was an error (Err result), the stream ends after it
    // (the StreamingHandler's contract: Err is terminal). The stream's own
    // end (None) triggers call.completed below.
}
// Natural stream end → write call.completed
let completed = EventEnvelope::completed(&request_id);
if let Err(err) = writer.write_frame(&completed).await {
    warn!(error = %err, "failed to write call.completed");
}

The ResponseEnvelope → EventEnvelope conversion (into()) already exists and produces call.responded for Ok and call.error for Err. The call.completed frame is written once when the stream ends naturally (not on error — an Err envelope is terminal, the stream ends after it, and we do NOT write call.completed after a call.error; the stream's None after an error is not a "natural end"). Track whether the last envelope was an error to decide whether to write call.completed. Alternatively, the StreamingHandler's contract is: Err ends the stream (the handler's stream yields the error then None), so after the loop, only write call.completed if the stream did not end on an error. Simplest correct approach: write call.completed only on natural end (the stream returned None without the last item being an Err). Track a last_was_error flag.

deadline: None for subscriptions

build_root_context sets deadline: Some(now + 30s). For the streaming path, the spec says deadline: None (unbounded — subscriptions are long-running). Set context.deadline = None after build_root_context for the streaming branch, or add a parameter to build_root_context. The deadline bounds the request/response call tree; a subscription has no such bound. Document this.

Abort cascade (ADR-016)

If call.aborted arrives for a streaming request ID, the stream is dropped (Rust Drop releases the handler's resources). The existing handle_abort path already removes the pending entry and cascades. For the streaming branch, the stream future being dropped (when the handle_stream task is cancelled or the call.aborted is processed) releases the handler's resources via Drop. No new abort code is needed — the existing handle_abort + the stream's Drop handle it. Verify the streaming pump task is cancellable (it's a tokio::spawn task; aborting the connection cancels it).

What this task does NOT do

• No client-side changes. The client CallConnection::subscribe() already works (it reads call.responded events until call.completed). This task is server-side only.
• No gateway changes. GatewayDispatch::invoke_streaming is http/gateway/invoke-streaming.

Acceptance Criteria

• dispatch_requested (or a new dispatch_requested_streaming) branches on op_type: Subscription → invoke_streaming(), Query/Mutation → invoke() (existing)
• handle_stream pumps the ResponseStream for the streaming branch: each ResponseEnvelope → EventEnvelope frame
• Natural stream end → call.completed frame written
• Err envelope → call.error frame written, stream ends after it (no call.completed after an error)
• deadline: None for the streaming branch (unbounded subscriptions)
• Abort: call.aborted for a streaming request drops the stream (Drop releases resources; existing handle_abort handles the pending entry)
• Existing Query/Mutation dispatch path unchanged (one call.responded/call.error frame, no call.completed)
• Unit test: Subscription op dispatch → multiple call.responded frames + call.completed on stream end
• Unit test: Subscription op handler yields Err → one call.error frame, no call.completed after
• Unit test: Query op dispatch unchanged (one frame, no call.completed)
• Unit test: call.aborted for streaming request → stream dropped
• cargo test -p alknet-call succeeds
• cargo clippy -p alknet-call --all-targets succeeds with no warnings
• cargo fmt --check -p alknet-call passes

References

• docs/architecture/decisions/049-streaming-handler-for-subscriptions.md — ADR-049 §6 (server-side dispatch branches on op_type)
• docs/architecture/crates/call/call-protocol.md — §CallAdapter Stream Handling (streaming branch: invoke_streaming → write each → call.completed; deadline: None; abort cascade)
• docs/architecture/decisions/016-abort-cascade-for-nested-calls.md — ADR-016 (stream drop on abort)

Notes

The streaming pump is a straightforward while let Some(envelope) = stream.next().await loop — not a complex abstraction. The tricky part is the call.completed semantics: write it on natural stream end, NOT after an Err (which is terminal). Track whether the last envelope was an error. deadline: None for subscriptions is a spec requirement — the 30s request/response deadline does not bound a long-running subscription. The abort cascade needs no new code: dropping the stream future (via task cancellation or handle_abort) releases the handler's resources through Rust's Drop. Pick the dispatch-entry shape (separate streaming method vs unified enum return) and document it — the spec frames it as a branch in handle_stream, so the branch should be visible there.

Summary

To be filled on completion