run.getReadable()/run.readable do not propagate cancel on disconnect, leaking stream listeners

[open-agents-bot]

Summary

When a workflow stream is returned directly from a route handler via run.readable or run.getReadable(), repeated client disconnects / refreshes do not appear to cancel the underlying workflow stream reader.

In local dev this eventually causes leaked EventEmitter listeners and MaxListenersExceededWarning warnings for chunk:* / close:* stream events.

This reproduces both when the stream is handed to createUIMessageStreamResponse(...) and when it is returned directly with new Response(run.getReadable(), ...).

Versions

• workflow: 4.1.0-beta.60
• @workflow/ai: 4.0.1-beta.54
• ai: 6.0.97
• @ai-sdk/react: 3.0.99
• next: 16.1.6

Minimal shape of the route

import { createUIMessageStreamResponse } from 'ai';
import { start } from 'workflow/api';

export async function POST() {
  const run = await start(myWorkflow, [/* args */]);
  return createUIMessageStreamResponse({
    stream: run.readable,
    headers: {
      'x-workflow-run-id': run.runId,
    },
  });
}

Resume route:

export async function GET(
  request: Request,
  { params }: { params: Promise<{ id: string }> },
) {
  const { id } = await params;
  const { searchParams } = new URL(request.url);
  const startIndex = Number.parseInt(searchParams.get('startIndex') ?? '0', 10);
  const run = getRun(id);

  return createUIMessageStreamResponse({
    stream: run.getReadable({ startIndex }),
  });
}

Repro

1. Start a workflow that streams for a while.
2. Connect to it from the browser.
3. Refresh repeatedly during the active stream, around 10 times is enough.
4. Keep resuming the same run.

Actual behavior

Eventually local dev logs show:

MaxListenersExceededWarning: Possible EventEmitter memory leak detected.
11 chunk:strm_..._user listeners added to [EventEmitter].
11 close:strm_..._user listeners added to [EventEmitter].

The app may still function, but old readers appear to remain subscribed.

Expected behavior

When the HTTP response is aborted because the client refreshes / disconnects, the abandoned workflow stream reader should be canceled and its listeners/subscriptions should be removed.

run.readable / run.getReadable() should be safe to hand directly to a streaming HTTP response without leaking listeners.

Likely root cause

• Run.readable just delegates to getReadable().
• getReadable() goes through the external reviver pipeline rather than returning the world stream directly.
• That pipeline creates a WorkflowServerReadableStream, but WorkflowServerReadableStream does not implement cancel().
• The reviver path then manually pumps that source stream through flushablePipe(...).
• When the outer HTTP response / consumer is aborted, cancellation does not seem to propagate all the way back to the source stream reader.
• In the local/postgres streamers, listener cleanup happens from the readable stream's own cancel(), so abandoned readers stay subscribed.

So the missing piece looks like cancel propagation through the built-in run.getReadable() pipeline, not just the route integration.

Workaround

I was able to avoid the leak by bypassing run.readable / run.getReadable() and instead:

1. reading the raw workflow stream directly
2. wrapping it in a ReadableStream whose cancel() calls reader.cancel()
3. deserializing it back into the expected chunk type

That strongly suggests the missing piece is cancel propagation in the built-in workflow readable.

Suggested fix

Propagate cancellation from the stream returned by run.readable / run.getReadable() all the way down to the underlying world stream reader.

That likely requires both:

1. implementing cancel(reason) on WorkflowServerReadableStream so it delegates to its inner reader, and
2. updating the flushablePipe(...) / reviver bridge so aborting the outer consumer cancels the source stream instead of only releasing locks.

Note

I also see ResponseAborted in Next dev logs during refresh, but that looks like expected framework disconnect noise. The real issue here is the accumulating MaxListenersExceededWarning.

[nicoalbanese]

Hope this is accurate 😊

[pranaygp]

fyi @TooTallNate this seems real and looks like there's already a PR for it #1354