diff --git a/README.md b/README.md
index 20cb1a4..2484729 100644
--- a/README.md
+++ b/README.md
@@ -103,6 +103,16 @@ plist, then remove it once the state directory has been populated.
   local sockets hanging and apps waiting on a dead link. When a probe fails,
   every connection on that forward is closed so clients see the drop and
   reconnect. Detection takes up to `--probe-interval` + `--probe-timeout`.
+- **Failures are reset, not closed.** How a connection ends is forwarded
+  faithfully. A target that closes cleanly gives the local client a FIN (an
+  ordinary EOF); a target that resets, errors, or disappears gives it an RST.
+  This matters for clients that hold idle connections: a FIN leaves the socket
+  writable, so a pooled client's *next* write still succeeds and only the one
+  after it fails, whereas an RST fails the very next read or write. It also
+  keeps a truncated response from looking like a complete one. The trade-off
+  is that an RST discards whatever was still queued in the send buffer — a
+  stream torn down this way was already incomplete, so flagging it beats
+  delivering a partial result that looks whole.
 - **Probes are real connections.** Each probe opens and immediately closes a
   TCP connection to the target. Chatty services may log these; raise
   `--probe-interval` to quiet them down, at the cost of slower detection.
diff --git a/main.go b/main.go
index 839a9e3..64753b2 100644
--- a/main.go
+++ b/main.go
@@ -254,8 +254,9 @@ func (p *proxy) markDown(cause error) {
 	if ln != nil {
 		ln.Close()
 	}
+	// RST, not FIN: these connections did not end, they broke. See abort.
 	for _, c := range conns {
-		c.Close()
+		abort(c)
 	}
 }
 
@@ -307,20 +308,46 @@ func (p *proxy) untrack(c net.Conn) {
 	delete(p.conns, c)
 }
 
+// abort closes c with a TCP RST instead of a FIN. A FIN says "the peer
+// finished normally", which is a lie when the target broke: an idle client
+// won't notice until its next write, and a client reading a response with no
+// declared length cannot tell a truncated body from a complete one. An RST
+// fails the peer's next read or write immediately and unambiguously.
+//
+// This discards anything still queued in the send buffer, which is the point:
+// a stream that ended this way was incomplete regardless, and flagging it
+// beats delivering a partial result that looks whole.
+func abort(c net.Conn) {
+	if tc, ok := c.(*net.TCPConn); ok {
+		tc.SetLinger(0)
+	}
+	c.Close()
+}
+
 func (p *proxy) handle(ctx context.Context, in net.Conn) {
 	defer p.untrack(in)
-	defer in.Close()
 
 	out, err := p.dial.Dial(ctx, "tcp", p.target)
 	if err != nil {
 		log.Printf("dial %s: %v", p.target, err)
 		p.nudge()
+		abort(in)
 		return
 	}
 	defer out.Close()
 
-	done := make(chan struct{}, 2)
-	go func() { io.Copy(out, in); done <- struct{}{} }()
-	go func() { io.Copy(in, out); done <- struct{}{} }()
-	<-done
+	// Which direction ended, and whether it ended cleanly. A clean EOF from the
+	// target is passed through as a FIN -- that is exactly what the client
+	// would have seen connecting directly. An error is passed through as an
+	// RST, for the same reason.
+	type result struct{ err error }
+	done := make(chan result, 2)
+	go func() { _, err := io.Copy(out, in); done <- result{err} }()
+	go func() { _, err := io.Copy(in, out); done <- result{err} }()
+
+	if r := <-done; r.err != nil {
+		abort(in)
+		return
+	}
+	in.Close()
 }
diff --git a/main_test.go b/main_test.go
index c43a961..4c23f42 100644
--- a/main_test.go
+++ b/main_test.go
@@ -6,6 +6,7 @@ import (
 	"io"
 	"net"
 	"sync"
+	"syscall"
 	"testing"
 	"time"
 )
@@ -19,6 +20,7 @@ type fakeTarget struct {
 
 	mu        sync.Mutex
 	reachable bool
+	accepted  []net.Conn
 }
 
 func newFakeTarget(t *testing.T) *fakeTarget {
@@ -35,12 +37,30 @@ func newFakeTarget(t *testing.T) *fakeTarget {
 			if err != nil {
 				return
 			}
+			ft.mu.Lock()
+			ft.accepted = append(ft.accepted, c)
+			ft.mu.Unlock()
 			go func() { io.Copy(c, c); c.Close() }()
 		}
 	}()
 	return ft
 }
 
+// killAccepted hard-closes every connection the target has accepted, as if the
+// service process were killed. reset picks RST (SO_LINGER 0) over a clean FIN.
+func (f *fakeTarget) killAccepted(reset bool) {
+	f.mu.Lock()
+	conns := f.accepted
+	f.accepted = nil
+	f.mu.Unlock()
+	for _, c := range conns {
+		if reset {
+			c.(*net.TCPConn).SetLinger(0)
+		}
+		c.Close()
+	}
+}
+
 func (f *fakeTarget) setReachable(v bool) {
 	f.mu.Lock()
 	defer f.mu.Unlock()
@@ -188,14 +208,18 @@ func TestTargetLossClosesLocalConnection(t *testing.T) {
 	// nothing but the probe loop can notice.
 	ft.setReachable(false)
 
-	// The client's blocking read must return, and reasonably promptly.
+	// The client's blocking read must return, promptly, and as a reset -- a
+	// plain EOF here would tell the client the stream ended normally.
 	c.SetReadDeadline(time.Now().Add(3 * time.Second))
-	if _, err := c.Read(buf); err == nil {
+	_, err = c.Read(buf)
+	if err == nil {
 		t.Fatal("read succeeded after target loss; want the connection closed")
-	} else if errors.Is(err, io.EOF) || isTimeout(err) {
-		if isTimeout(err) {
-			t.Fatal("read blocked after target loss; local socket was never closed")
-		}
+	}
+	if isTimeout(err) {
+		t.Fatal("read blocked after target loss; local socket was never closed")
+	}
+	if !isReset(err) {
+		t.Errorf("read err = %v, want a connection reset (a clean EOF would falsely signal a complete stream)", err)
 	}
 
 	// New connects must be refused too.
@@ -206,3 +230,74 @@ func isTimeout(err error) bool {
 	var ne net.Error
 	return errors.As(err, &ne) && ne.Timeout()
 }
+
+func isReset(err error) bool {
+	return errors.Is(err, syscall.ECONNRESET)
+}
+
+// A target process that dies drops its connections with a FIN or an RST. The
+// local client must see that within a round trip -- it must NOT have to wait
+// for the reachability probe, which is the slow fallback for a peer that
+// vanishes without any TCP signal at all.
+func TestTargetHardCloseIsNoticedImmediately(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		reset bool
+	}{
+		{"fin", false},
+		{"rst", true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			ft := newFakeTarget(t)
+			local := freePort(t)
+			// Probe interval far longer than the assertion window, so a pass
+			// can only mean the per-connection path noticed.
+			p := &proxy{
+				dial:     ft,
+				local:    local,
+				target:   "target:1234",
+				interval: 30 * time.Second,
+				timeout:  time.Second,
+				recheck:  make(chan struct{}, 1),
+				conns:    make(map[net.Conn]struct{}),
+			}
+			ctx, cancel := context.WithCancel(context.Background())
+			done := make(chan struct{})
+			go func() { defer close(done); p.run(ctx) }()
+			defer func() { cancel(); <-done }()
+
+			waitFor(t, "local port to accept", func() bool { return !localPortRefused(local) })
+
+			c, err := net.Dial("tcp", local)
+			if err != nil {
+				t.Fatalf("dial local: %v", err)
+			}
+			defer c.Close()
+			if _, err := c.Write([]byte("ping")); err != nil {
+				t.Fatalf("write: %v", err)
+			}
+			buf := make([]byte, 4)
+			c.SetReadDeadline(time.Now().Add(3 * time.Second))
+			if _, err := io.ReadFull(c, buf); err != nil {
+				t.Fatalf("read echo: %v", err)
+			}
+
+			start := time.Now()
+			ft.killAccepted(tc.reset)
+
+			c.SetReadDeadline(time.Now().Add(5 * time.Second))
+			_, err = c.Read(buf)
+			elapsed := time.Since(start)
+			if err == nil {
+				t.Fatal("read succeeded after target hard-close")
+			}
+			if isTimeout(err) {
+				t.Fatalf("client never noticed target hard-close (blocked %s)", elapsed)
+			}
+			t.Logf("client noticed after %s (err: %v)", elapsed, err)
+			if elapsed > time.Second {
+				t.Errorf("client took %s to notice; want sub-second", elapsed)
+			}
+		})
+	}
+}