diff --git a/engine/iouring/async_write_order_test.go b/engine/iouring/async_write_order_test.go
new file mode 100644
index 00000000..363be68e
--- /dev/null
+++ b/engine/iouring/async_write_order_test.go
@@ -0,0 +1,247 @@
+//go:build linux
+
+package iouring
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"io"
+	"net/http"
+	"strconv"
+	"testing"
+	"time"
+
+	"golang.org/x/sys/unix"
+
+	"github.com/goceleris/celeris/internal/conn"
+	"github.com/goceleris/celeris/protocol/h2/stream"
+)
+
+// celeris#751: with AsyncHandlers, the dispatch goroutine writes each response
+// straight to the socket (unix.Write at the end of a runAsyncHandler
+// iteration), and it did so without asking whether a ring SEND of the
+// connection's earlier bytes was still in flight. Those bytes are the rest of
+// the previous response: its direct write was short, the goroutine left the
+// remainder in writeBuf, and the worker moved it to sendBuf and submitted a
+// SEND. A pipelined request answered while that SEND is outstanding went out
+// ahead of it, in the middle of the previous response. The h2c-upgrade exit's
+// direct write of the 101 had the same gap. The detached conns' guarded
+// writeFn has always refused the raw write while a ring SEND is outstanding.
+//
+// The test plays the kernel and the client on one conn of an fdlFixture: it
+// captures the SEND the worker submits, performs it itself (writing sendBuf to
+// the socket) when it chooses, and reads the client end between steps. The
+// window is therefore an input: the client has drained the socket, so a
+// direct write issued while the SEND is outstanding is admitted.
+
+// orderBody751 is the first response's body: far larger than a socketpair's
+// buffer, so the goroutine's direct write is short and the rest becomes a
+// ring SEND. A repeating pattern, so a foreign byte sequence inside it is seen.
+var orderBody751 = bytes.Repeat([]byte("0123456789abcdef"), (1<<20)/16)
+
+type orderHandler751 struct{}
+
+func (orderHandler751) HandleStream(_ context.Context, s *stream.Stream) error {
+	if s.ResponseWriter == nil {
+		return nil
+	}
+	body := []byte("ok")
+	if s.Path == "/big" {
+		body = orderBody751
+	}
+	return s.ResponseWriter.WriteResponse(s, 200,
+		[][2]string{{"content-type", "text/plain"}, {"content-length", strconv.Itoa(len(body))}}, body)
+}
+
+// orderClient751 is the client end of the fixture's socketpair, read
+// non-blocking, so the test decides when the client reads.
+type orderClient751 struct {
+	t    *testing.T
+	peer int
+	got  []byte
+}
+
+// drain reads everything the socket holds now.
+func (c *orderClient751) drain() int {
+	c.t.Helper()
+	n0 := len(c.got)
+	var b [64 << 10]byte
+	for {
+		n, err := unix.Read(c.peer, b[:])
+		if n > 0 {
+			c.got = append(c.got, b[:n]...)
+			continue
+		}
+		if err == unix.EAGAIN || err == unix.EINTR {
+			return len(c.got) - n0
+		}
+		c.t.Fatalf("client read: %d, %v", n, err)
+	}
+}
+
+// waitDispatchIdle waits until f's dispatch goroutine has run everything
+// delivered to it and reached its park (or, with exited, has exited). Its
+// write and its enqueue happen before either, so both are done on return.
+func waitDispatchIdle751(t *testing.T, f *fdlFixture, exited bool) {
+	t.Helper()
+	for deadline := time.Now().Add(10 * time.Second); ; time.Sleep(time.Millisecond) {
+		f.cs.asyncInMu.Lock()
+		idle := len(f.cs.asyncInBuf) == 0 && (f.cs.asyncRun && f.cs.asyncParked || exited && !f.cs.asyncRun)
+		f.cs.asyncInMu.Unlock()
+		if idle {
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatal("the dispatch goroutine did not finish its requests within 10s")
+		}
+	}
+}
+
+// kernelSend performs the SEND the worker has in flight the way the kernel
+// does, writing cs.sendBuf to the socket (the client reads whenever the
+// socket is full), and completes it. It then runs the worker's per-iteration
+// work (the detach queue, the dirty list) and repeats while a SEND is in
+// flight. It reports how many SENDs it performed.
+func kernelSend751(t *testing.T, f *fdlFixture, c *orderClient751) int {
+	t.Helper()
+	sends := 0
+	for {
+		f.w.drainDetachQueue()
+		f.w.flushDirty()
+		for _, s := range takeSQEs(f.w.ring) {
+			if s.op != opSEND && s.op != opRECV {
+				t.Fatalf("the worker placed %v, want only SENDs and recv arms", s)
+			}
+		}
+		if !f.cs.sending {
+			return sends
+		}
+		if sends++; sends > 16 {
+			t.Fatal("more than 16 SENDs for two responses")
+		}
+		for b := f.cs.sendBuf; len(b) > 0; {
+			n, err := unix.Write(f.fd, b)
+			if n > 0 {
+				b = b[n:]
+				continue
+			}
+			if err != unix.EAGAIN && err != unix.EINTR {
+				t.Fatalf("kernel send: %v", err)
+			}
+			if c.drain() == 0 {
+				t.Fatal("kernel send: the socket is full and the client has nothing to read")
+			}
+		}
+		f.process(f.sendCQE())
+	}
+}
+
+// TestAsyncResponseWaitsForAnInFlightRingSend: the second of two pipelined
+// requests on a promoted async conn is answered while the ring SEND of the
+// first response's tail is in flight. The client must receive the first
+// response whole, then the second: arm response is a plain response, arm
+// h2c_upgrade is the 101 of an h2c upgrade.
+func TestAsyncResponseWaitsForAnInFlightRingSend(t *testing.T) {
+	for _, arm := range []string{"response", "h2c_upgrade"} {
+		t.Run(arm, func(t *testing.T) {
+			f := newFDLFixture(t, true)
+			f.w.handler = orderHandler751{}
+			if arm == "h2c_upgrade" {
+				f.w.cfg.EnableH2Upgrade = true
+				f.w.h2cfg = conn.H2Config{MaxConcurrentStreams: 100, InitialWindowSize: 65535,
+					MaxFrameSize: 16384, MaxRequestBodySize: 1 << 20}
+				f.w.initProtocol(f.cs) // a fresh h1State with the upgrade on, as onAcceptedFD makes it
+			}
+			f.cs.asyncPromoted.Store(true)
+			f.armFirstRecv()
+			if err := unix.SetNonblock(f.peer, true); err != nil {
+				t.Fatalf("nonblock: %v", err)
+			}
+			client := &orderClient751{t: t, peer: f.peer}
+			t.Cleanup(func() {
+				f.cs.asyncInMu.Lock()
+				f.cs.asyncClosed.Store(true)
+				f.cs.asyncCond.Broadcast()
+				f.cs.asyncInMu.Unlock()
+				f.w.asyncWG.Wait()
+			})
+
+			// Request 1: the goroutine's direct write fills the socket, and the
+			// worker submits the rest as a ring SEND, which stays in flight.
+			f.deliver("GET /big HTTP/1.1\r\nHost: x\r\n\r\n")
+			waitDispatchIdle751(t, f, false)
+			f.w.drainDetachQueue()
+			f.w.flushDirty()
+			var placed []sqeRec
+			for _, s := range takeSQEs(f.w.ring) {
+				if s.op != opRECV {
+					placed = append(placed, s)
+				}
+			}
+			if len(placed) != 1 || placed[0].op != opSEND || !f.cs.sending || len(f.cs.sendBuf) == 0 {
+				t.Fatalf("apparatus: after request 1 the worker placed %v (sending=%v, %d bytes to send), want one "+
+					"SEND of the response's tail", placed, f.cs.sending, len(f.cs.sendBuf))
+			}
+			tail := len(f.cs.sendBuf)
+
+			// The client reads what the direct write delivered: the socket is empty
+			// again, and the SEND is still the kernel's.
+			head := client.drain()
+
+			// Request 2, answered while the SEND is outstanding.
+			req2 := "GET /small HTTP/1.1\r\nHost: x\r\n\r\n"
+			if arm == "h2c_upgrade" {
+				req2 = h2cUpgradeHead722("GET", 0)
+			}
+			f.deliver(req2)
+			waitDispatchIdle751(t, f, arm == "h2c_upgrade")
+			early := client.drain()
+
+			// The kernel completes the SEND (and any the worker submits after it).
+			sends := kernelSend751(t, f, client)
+			client.drain()
+			t.Logf("celeris751 ORDER arm=%s head=%d tail=%d early=%d sends=%d total=%d",
+				arm, head, tail, early, sends, len(client.got))
+			if early != 0 {
+				t.Errorf("celeris#751: %d bytes of the answer to request 2 reached the client while the ring SEND "+
+					"of response 1's last %d bytes was still in flight", early, tail)
+			}
+
+			br := bufio.NewReader(bytes.NewReader(client.got))
+			r1, err := http.ReadResponse(br, nil)
+			if err != nil {
+				t.Fatalf("response 1: %v", err)
+			}
+			b1, err := io.ReadAll(r1.Body)
+			if err != nil || !bytes.Equal(b1, orderBody751) {
+				at := 0
+				for at < min(len(b1), len(orderBody751)) && b1[at] == orderBody751[at] {
+					at++
+				}
+				t.Fatalf("celeris#751: response 1's body is not intact (%d bytes, err %v; first difference at "+
+					"byte %d: %q): the next response was written into it", len(b1), err, at,
+					b1[at:min(at+48, len(b1))])
+			}
+			r2, err := http.ReadResponse(br, nil)
+			if err != nil {
+				t.Fatalf("celeris#751: the answer to request 2 does not follow response 1: %v", err)
+			}
+			want := 200
+			if arm == "h2c_upgrade" {
+				want = http.StatusSwitchingProtocols
+			}
+			if r2.StatusCode != want {
+				t.Errorf("request 2 got status %d, want %d", r2.StatusCode, want)
+			}
+			if arm == "response" {
+				if b2, err := io.ReadAll(r2.Body); err != nil || string(b2) != "ok" {
+					t.Errorf("response 2 body %q, err %v; want \"ok\"", b2, err)
+				}
+				if rest, _ := io.ReadAll(br); len(rest) != 0 {
+					t.Errorf("%d bytes after response 2: %q", len(rest), rest[:min(48, len(rest))])
+				}
+			}
+		})
+	}
+}
diff --git a/engine/iouring/worker.go b/engine/iouring/worker.go
index 0be62195..f9e079f7 100644
--- a/engine/iouring/worker.go
+++ b/engine/iouring/worker.go
@@ -4743,8 +4743,11 @@ func (w *Worker) runAsyncHandler(cs *connState) {
 			// 101 to whatever real fd holds that number (celeris#538).
 			// Leaving the bytes in writeBuf takes the same disposition as
 			// the EAGAIN branch below — the worker ring-sends them, and the
-			// ring resolves the index correctly.
-			if promoteErr == nil && !cs.fixedFile && len(cs.writeBuf) > 0 {
+			// ring resolves the index correctly. Nor while a ring SEND of
+			// the conn's earlier bytes is outstanding (celeris#751): the 101
+			// would reach the client ahead of the rest of the previous
+			// response. The worker sends writeBuf after that SEND completes.
+			if promoteErr == nil && !cs.fixedFile && len(cs.writeBuf) > 0 && !ringSendOutstanding(cs) {
 				n, werr := unix.Write(cs.fd, cs.writeBuf)
 				switch {
 				case werr == nil && n == len(cs.writeBuf):
@@ -4835,6 +4838,14 @@ func (w *Worker) runAsyncHandler(cs *connState) {
 			// full socket buffer — hand the bytes to the worker, whose ring
 			// SEND resolves the index correctly.
 			partial = true
+		} else if processErr == nil && len(cs.writeBuf) > 0 && ringSendOutstanding(cs) {
+			// A ring SEND of this conn's earlier bytes is outstanding: the
+			// rest of a previous response whose own direct write was short
+			// (celeris#751). Writing now would put these bytes on the wire
+			// ahead of it, inside the previous response of a pipelining
+			// client. Leave them in writeBuf, as for a full socket buffer:
+			// the worker's completeSend sends writeBuf after that SEND.
+			partial = true
 		} else if processErr == nil && len(cs.writeBuf) > 0 {
 			n, werr := unix.Write(cs.fd, cs.writeBuf)
 			if werr != nil {
@@ -4883,6 +4894,17 @@ func (w *Worker) runAsyncHandler(cs *connState) {
 	}
 }
 
+// ringSendOutstanding reports whether a ring SEND of cs's earlier bytes is in
+// flight, or owed, so that a raw unix.Write of writeBuf now would reach the
+// wire ahead of them (celeris#751): a SEND or its SEND_ZC notification is
+// outstanding, or sendBuf/bodyBuf hold bytes the worker has taken from
+// writeBuf and not yet sent. The worker mutates all four under cs.detachMu,
+// which the caller holds, so no SEND can start while it does. The detached
+// conns' guarded writeFn tests the same condition.
+func ringSendOutstanding(cs *connState) bool {
+	return cs.sending || cs.zcNotifPending || len(cs.sendBuf) > 0 || len(cs.bodyBuf) > 0
+}
+
 func (w *Worker) makeWriteFn(cs *connState) func([]byte) {
 	return func(data []byte) {
 		if cs.closing {