diff --git a/wsflate/writer.go b/wsflate/writer.go
index 0342ccb..64e4a48 100644
--- a/wsflate/writer.go
+++ b/wsflate/writer.go
@@ -54,7 +54,10 @@ type Writer struct {
 	ctor func(io.Writer) Compressor
 	c    Compressor
 	cbuf cbuf
-	err  error
+	// tailChecked reports whether checkTail() already consumed the stream
+	// tail for the current (non-reset) message.
+	tailChecked bool
+	err         error
 }
 
 // NewWriter returns a new Writer.
@@ -74,6 +77,7 @@ func NewWriter(w io.Writer, ctor func(io.Writer) Compressor) *Writer {
 // Any not flushed data will be lost.
 func (w *Writer) Reset(dest io.Writer) {
 	w.err = nil
+	w.tailChecked = false
 	w.cbuf.reset(dest)
 	if x, ok := w.c.(WriteResetter); ok {
 		x.Reset(&w.cbuf)
@@ -119,11 +123,53 @@ func (w *Writer) Err() error {
 	return w.err
 }
 
+// tailBFINALFixed is the last 4 bytes cbuf holds when a compressor ends the
+// stream with an empty final fixed-Huffman block (BFINAL=1, BTYPE=01) instead
+// of the traditional empty non-final stored block.
+//
+// Starting with Go 1.27 (golang/go commit 388145c, "compress/flate: improve
+// compression speed"), compress/flate finishes a stream with a 3-byte empty
+// final fixed block (0x03 0x00 0x00: the 0x03 header has BFINAL|BTYPE=01 and
+// 5 zero padding bits) instead of the 4-byte empty stored block (0x00 0x00
+// 0xff 0xff) used before. cbuf holds back the last 4 bytes, so the new
+// tail reads as {0xff, 0xff, 0x03, 0x00}.
+//
+// Both tails are valid per RFC 7692 section 7.2.3.4 ("Using a DEFLATE Block
+// with BFINAL Set to 1"), which allows ending a message on a BFINAL block
+// rather than on an empty non-final stored block.
+var tailBFINALFixed = [4]byte{0xff, 0xff, 0x03, 0x00}
+
 func (w *Writer) checkTail() {
-	if w.err == nil && w.cbuf.buf != compressionTail {
+	if w.err != nil {
+		return
+	}
+	switch {
+	case w.tailChecked:
+		// Tail was already consumed by a previous checkTail() call (e.g.
+		// Flush() followed by Close()). Nothing left to do.
+	case w.cbuf.buf == compressionTail:
+		// Classic tail: the compressor ended on an empty non-final stored
+		// block. cbuf's 4 held bytes are the 0x00 0x00 0xff 0xff that the
+		// receiver appends back per RFC 7692 section 7.2.2, so they are
+		// dropped (nothing is written to dest).
+	case w.cbuf.buf == tailBFINALFixed:
+		// Go 1.27 tail: the stream ends with an empty final fixed-Huffman
+		// block (0x03 0x00) right after the empty non-final stored block.
+		// cbuf has already flushed the first 3 bytes of that stored block,
+		// so pass all 4 held bytes (FF FF 03 00) through to the destination
+		// to reconstruct the full, self-terminating stream:
+		//   … [00 00 00 FF FF] [03 00]
+		// (empty stored block + final fixed block). The receiver then
+		// appends 0x00 0x00 0xff 0xff per RFC 7692 7.2.2, which lands after
+		// the BFINAL block and is simply ignored by the decompressor.
+		w.cbuf.flush(w.cbuf.buf[:])
+		w.cbuf.buf = [4]byte{}
+		w.cbuf.n = 0
+		w.tailChecked = true
+	default:
 		w.err = fmt.Errorf(
-			"wsflate: bad compressor: unexpected stream tail: %#x vs %#x",
-			w.cbuf.buf, compressionTail,
+			"wsflate: bad compressor: unexpected stream tail: %#x (want %#x or %#x)",
+			w.cbuf.buf, compressionTail, tailBFINALFixed,
 		)
 	}
 }
diff --git a/wsflate/writer_tail_test.go b/wsflate/writer_tail_test.go
new file mode 100644
index 0000000..ef61a37
--- /dev/null
+++ b/wsflate/writer_tail_test.go
@@ -0,0 +1,147 @@
+package wsflate
+
+import (
+	"bytes"
+	"compress/flate"
+	"io"
+	"testing"
+)
+
+// blockCompressor is a Compressor that emits an exact, pre-recorded byte
+// stream (the blocks, in order) then writes any literal Write() input
+// through unchanged. It lets tests pin the exact tail bytes independent of
+// the Go toolchain's compress/flate encoder.
+type blockCompressor struct {
+	dst    io.Writer
+	blocks [][]byte
+	pos    int
+}
+
+func (f *blockCompressor) emit() {
+	for f.pos < len(f.blocks) {
+		f.dst.Write(f.blocks[f.pos])
+		f.pos++
+	}
+}
+
+func (f *blockCompressor) Write(p []byte) (int, error) {
+	f.emit()
+	f.dst.Write(p)
+	return len(p), nil
+}
+
+func (f *blockCompressor) Flush() error { f.emit(); return nil }
+
+func (f *blockCompressor) Close() error { f.emit(); return nil }
+
+// TestCheckTailClassic pins the legacy behaviour: a compressor that ends on
+// an empty non-final stored block produces a 0x00 0x00 0xff 0xff held tail,
+// which must be dropped (nothing extra written to the destination) and must
+// not raise an error.
+func TestCheckTailClassic(t *testing.T) {
+	// Stream: an empty final fixed block (03 00) followed by an empty
+	// non-final stored block (00 00 00 FF FF). The last 4 bytes held by
+	// cbuf are therefore 00 00 FF FF (compressionTail).
+	var dest bytes.Buffer
+	w := NewWriter(&dest, func(w io.Writer) Compressor {
+		return &blockCompressor{
+			dst:    w,
+			blocks: [][]byte{{0x03, 0x00}, {0x00, 0x00, 0x00, 0xff, 0xff}},
+		}
+	})
+	w.Write(nil)
+	if err := w.Flush(); err != nil {
+		t.Fatalf("Flush() error: %v", err)
+	}
+	// Flush() then Close() (as CompressTo does) must not double-handle the tail.
+	if err := w.Close(); err != nil {
+		t.Fatalf("Close() error: %v", err)
+	}
+	if err := w.Err(); err != nil {
+		t.Fatalf("Writer.Err() = %v; want nil", err)
+	}
+	// cbuf already flushed the first 3 bytes of the stored block; the final
+	// 4 (00 00 ff ff) must be dropped, leaving dest = 03 00 00.
+	got := dest.Bytes()
+	if want := []byte{0x03, 0x00, 0x00}; !bytes.Equal(got, want) {
+		t.Fatalf("dest = % X; want % X", got, want)
+	}
+}
+
+// TestCheckTailBFINALFixed pins the Go 1.27 behaviour: a compressor that ends
+// on an empty final fixed-Huffman block (BFINAL=1, BTYPE=01) produces a
+// FF FF 03 00 held tail. checkTail must pass all 4 held bytes through so the
+// on-wire payload is a complete, self-terminating DEFLATE stream, and must
+// not raise an error. This branch is exercised here on every toolchain, not
+// only on Go 1.27.
+func TestCheckTailBFINALFixed(t *testing.T) {
+	// Stream ends: ... [00 00 00 FF FF] [03 00]
+	//   (empty non-final stored block)(empty final fixed block)
+	// cbuf holds the last 4 bytes: FF FF 03 00 (tailBFINALFixed).
+	var dest bytes.Buffer
+	w := NewWriter(&dest, func(w io.Writer) Compressor {
+		return &blockCompressor{
+			dst:    w,
+			blocks: [][]byte{{0x03, 0x00}, {0x00, 0x00, 0x00, 0xff, 0xff}, {0x03, 0x00}},
+		}
+	})
+	w.Write(nil)
+	if err := w.Flush(); err != nil {
+		t.Fatalf("Flush() error: %v", err)
+	}
+	if err := w.Close(); err != nil {
+		t.Fatalf("Close() error: %v", err)
+	}
+	if err := w.Err(); err != nil {
+		t.Fatalf("Writer.Err() = %v; want nil", err)
+	}
+	// Full reconstructed stream: 03 00 00 00 00 FF FF 03 00
+	// (9 bytes). The 03 00 final block makes it self-terminating.
+	got := dest.Bytes()
+	if want := []byte{0x03, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0x03, 0x00}; !bytes.Equal(got, want) {
+		t.Fatalf("dest = % X; want % X", got, want)
+	}
+	// The on-wire payload must round-trip through an RFC 7692 receiver
+	// (append 00 00 FF FF, decompress).
+	var out bytes.Buffer
+	fr := flate.NewReader(io.MultiReader(
+		bytes.NewReader(got), bytes.NewReader([]byte{0, 0, 0xff, 0xff}),
+	))
+	if _, err := io.Copy(&out, fr); err != nil {
+		t.Fatalf("decompressing payload: %v", err)
+	}
+}
+
+// TestCheckTailUnexpected ensures tails matching neither shape are still
+// rejected, preserving the original validation.
+func TestCheckTailUnexpected(t *testing.T) {
+	var dest bytes.Buffer
+	w := NewWriter(&dest, func(w io.Writer) Compressor {
+		return &blockCompressor{
+			dst:    w,
+			blocks: [][]byte{{0xaa, 0xbb, 0xcc, 0xdd}},
+		}
+	})
+	w.Write(nil)
+	w.Flush()
+	if w.Err() == nil {
+		t.Fatal("Writer.Err() = nil; want non-nil for unexpected tail")
+	}
+}
+
+// TestWriterRoundTripDefault is the end-to-end guard that the default
+// compress/flate based helper still round-trips on the current toolchain.
+func TestWriterRoundTripDefault(t *testing.T) {
+	data := []byte("hello, flate!")
+	var compressed bytes.Buffer
+	if err := DefaultHelper.CompressTo(&compressed, data); err != nil {
+		t.Fatalf("CompressTo() error: %v", err)
+	}
+	var out bytes.Buffer
+	if err := DefaultHelper.DecompressTo(&out, compressed.Bytes()); err != nil {
+		t.Fatalf("DecompressTo() error: %v", err)
+	}
+	if !bytes.Equal(out.Bytes(), data) {
+		t.Fatalf("round-trip mismatch: got %q; want %q", out.String(), string(data))
+	}
+}
diff --git a/wsflate/zz_edge_test.go b/wsflate/zz_edge_test.go
new file mode 100644
index 0000000..6013aaf
--- /dev/null
+++ b/wsflate/zz_edge_test.go
@@ -0,0 +1,135 @@
+package wsflate
+
+import (
+	"bytes"
+	"compress/flate"
+	"io"
+	"strings"
+	"testing"
+)
+
+func TestEdgeCasesRoundTrip(t *testing.T) {
+	cases := map[string][]byte{
+		"empty":    {},
+		"short":    []byte("hi"),
+		"one-byte": []byte("a"),
+		"exact-4":  []byte("abcd"),
+		"big-text": []byte(strings.Repeat("The quick brown fox. ", 1000)),
+	}
+	cases["randomish"] = []byte(strings.Repeat("random-lookin data | structure", 300))
+	for name, data := range cases {
+		t.Run(name, func(t *testing.T) {
+			var buf bytes.Buffer
+			w := NewWriter(&buf, func(w io.Writer) Compressor { f, _ := flate.NewWriter(w, 9); return f })
+			w.Write(data)
+			w.Flush()
+			w.Close()
+			if err := w.Err(); err != nil {
+				t.Fatalf("Writer.Err(): %v", err)
+			}
+			var out bytes.Buffer
+			r := NewReader(&buf, func(r io.Reader) Decompressor { return flate.NewReader(r) })
+			io.Copy(&out, r)
+			if err := r.Err(); err != nil {
+				t.Fatalf("Reader.Err(): %v (payload=% X)", err, buf.Bytes())
+			}
+			if !bytes.Equal(out.Bytes(), data) {
+				t.Fatalf("mismatch: got %d bytes want %d", out.Len(), len(data))
+			}
+		})
+	}
+}
+
+// Each message self-contained (DefaultParameters / no context takeover), with
+// fragmented writes and mid-message Flush() calls.
+func TestEdgeNoContextTakeover(t *testing.T) {
+	msgs := []string{
+		"first message with some content",
+		"",
+		"second message " + strings.Repeat("structure ", 40),
+		"tiny",
+		"abcd",
+	}
+	for i, m := range msgs {
+		var buf bytes.Buffer
+		w := NewWriter(&buf, func(w io.Writer) Compressor { f, _ := flate.NewWriter(w, 9); return f })
+		data := []byte(m)
+		const chunk = 5
+		for len(data) > 0 {
+			n := chunk
+			if n > len(data) {
+				n = len(data)
+			}
+			w.Write(data[:n])
+			data = data[n:]
+			if len(data)%2 == 0 {
+				if err := w.Flush(); err != nil {
+					t.Fatalf("msg %d mid-flush: %v", i, err)
+				}
+			}
+		}
+		w.Flush()
+		w.Close()
+		if err := w.Err(); err != nil {
+			t.Fatalf("msg %d writer err: %v (payload=% X)", i, err, buf.Bytes())
+		}
+		var out bytes.Buffer
+		r := NewReader(&buf, func(r io.Reader) Decompressor { return flate.NewReader(r) })
+		io.Copy(&out, r)
+		if err := r.Err(); err != nil {
+			t.Fatalf("msg %d reader err: %v (payload=% X)", i, err, buf.Bytes())
+		}
+		if out.String() != m {
+			t.Fatalf("msg %d: got %q want %q", i, out.String(), m)
+		}
+	}
+}
+
+// Context takeover: ONE long-lived writer. Each message ends on Flush(); the
+// final one on Close(). Per RFC 7692 the per-message payloads are byte-aligned
+// continuations of a single stream, so we reconstruct each message's payload
+// from the wire offsets and feed all payloads (each + 00 00 FF FF) to ONE
+// long-lived reader with a carried-over window.
+func TestEdgeContextTakeover(t *testing.T) {
+	msgs := []string{
+		"first message establishes the window with repeatable content ",
+		"now this message should reference the previous window content ",
+		"and yet another referencing the window ",
+		"shorter",
+	}
+	var wire bytes.Buffer
+	w := NewWriter(&wire, func(w io.Writer) Compressor { f, _ := flate.NewWriter(w, 9); return f })
+	for i, m := range msgs {
+		if _, err := w.Write([]byte(m)); err != nil {
+			t.Fatalf("msg %d write: %v", i, err)
+		}
+		if err := w.Flush(); err != nil {
+			t.Fatalf("msg %d flush: %v", i, err)
+		}
+		if err := w.Err(); err != nil {
+			t.Fatalf("msg %d writer err: %v", i, err)
+		}
+	}
+	if err := w.Close(); err != nil {
+		t.Fatalf("close: %v", err)
+	}
+	if err := w.Err(); err != nil {
+		t.Fatalf("final writer err: %v (wire=% X)", err, wire.Bytes())
+	}
+
+	// Context takeover = one continuous byte-aligned stream, terminated by the
+	// final BFINAL block. Feed the whole wire buffer to a single long-lived
+	// reader (window carried); no per-message suffix is added.
+	var out bytes.Buffer
+	r := NewReader(&wire, func(r io.Reader) Decompressor { return flate.NewReader(r) })
+	if _, err := io.Copy(&out, r); err != nil {
+		t.Fatalf("read: %v (wire=% X)", err, wire.Bytes())
+	}
+	if err := r.Err(); err != nil {
+		t.Fatalf("reader err: %v", err)
+	}
+	want := strings.Join(msgs, "")
+	if out.String() != want {
+		t.Fatalf("context-takeover concatenation mismatch:\n got %q\nwant %q", out.String(), want)
+	}
+}