diff --git a/internal/delivery/voicesink/voicesink_test.go b/internal/delivery/voicesink/voicesink_test.go
new file mode 100644
index 0000000..568a560
--- /dev/null
+++ b/internal/delivery/voicesink/voicesink_test.go
@@ -0,0 +1,218 @@
+package voicesink
+
+import (
+	"context"
+	"encoding/binary"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/audio"
+	"github.com/kami/maven/internal/delivery"
+	"github.com/kami/maven/internal/tts"
+	"github.com/kami/maven/internal/voice"
+)
+
+type frameResult struct {
+	buf []byte
+	err error
+}
+
+func readFrame(r net.Conn, ch chan<- frameResult) {
+	var hdr [4]byte
+	if _, err := io.ReadFull(r, hdr[:]); err != nil {
+		ch <- frameResult{err: err}
+		return
+	}
+	n := binary.BigEndian.Uint32(hdr[:])
+	buf := make([]byte, n)
+	if _, err := io.ReadFull(r, buf); err != nil {
+		ch <- frameResult{err: err}
+		return
+	}
+	ch <- frameResult{buf: buf}
+}
+
+func unmarshalFrame(t *testing.T, buf []byte) (kind string, params json.RawMessage) {
+	t.Helper()
+	var frame struct {
+		Kind   string          `json:"kind"`
+		Params json.RawMessage `json:"p"`
+	}
+	if err := json.Unmarshal(buf, &frame); err != nil {
+		t.Fatalf("unmarshal frame: %v", err)
+	}
+	return frame.Kind, frame.Params
+}
+
+func unmarshalPush(t *testing.T, raw json.RawMessage) voice.AudioNudgePush {
+	t.Helper()
+	var push voice.AudioNudgePush
+	if err := json.Unmarshal(raw, &push); err != nil {
+		t.Fatalf("unmarshal push params: %v", err)
+	}
+	return push
+}
+
+// ---- nil guards ----
+
+func TestSinkSend_NilTTSError(t *testing.T) {
+	s := New(nil, voice.NewSessions())
+	err := s.Send(context.Background(), delivery.Sendable{Body: "test"})
+	if err == nil || !strings.Contains(err.Error(), "tts synthesizer not wired") {
+		t.Fatalf("expected tts-not-wired error, got %v", err)
+	}
+}
+
+func TestSinkSend_NilSessionsError(t *testing.T) {
+	s := New(tts.NewStub(), nil)
+	err := s.Send(context.Background(), delivery.Sendable{Body: "test"})
+	if err == nil || !strings.Contains(err.Error(), "sessions registry not wired") {
+		t.Fatalf("expected sessions-not-wired error, got %v", err)
+	}
+}
+
+// ---- no session ----
+
+func TestSinkSend_NoSessionReturnsErrVoiceNoSession(t *testing.T) {
+	sess := voice.NewSessions()
+	s := New(tts.NewStub(), sess)
+	err := s.Send(context.Background(), delivery.Sendable{Body: "hello"})
+	if !errors.Is(err, delivery.ErrVoiceNoSession) {
+		t.Fatalf("expected ErrVoiceNoSession, got %v", err)
+	}
+}
+
+// ---- successful push ----
+
+func TestSinkSend_WithSessionPushesAudio(t *testing.T) {
+	r, w := net.Pipe()
+	defer r.Close()
+	defer w.Close()
+
+	ch := make(chan frameResult, 1)
+	go readFrame(r, ch)
+
+	ctx := context.Background()
+	sess := voice.NewSessions()
+	sess.Add(w, voice.SurfacePCClient)
+	s := New(tts.NewStub(), sess)
+
+	err := s.Send(ctx, delivery.Sendable{
+		RuleName: "test_rule",
+		Body:     "Hello world",
+		Kind:     delivery.KindNudge,
+		Ts:       time.Now(),
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	res := <-ch
+	if res.err != nil {
+		t.Fatalf("read frame: %v", res.err)
+	}
+
+	kind, params := unmarshalFrame(t, res.buf)
+	if kind != "audio_nudge" {
+		t.Errorf("push kind = %q, want %q", kind, "audio_nudge")
+	}
+
+	push := unmarshalPush(t, params)
+	if push.RuleName != "test_rule" {
+		t.Errorf("rule_name = %q, want %q", push.RuleName, "test_rule")
+	}
+	if push.Text != "Hello world" {
+		t.Errorf("text = %q, want %q", push.Text, "Hello world")
+	}
+	if push.Audio.Format != audio.PCM16kMono {
+		t.Errorf("audio format = %v, want %v", push.Audio.Format, audio.PCM16kMono)
+	}
+	if len(push.Audio.Bytes) == 0 {
+		t.Error("audio bytes are empty")
+	}
+}
+
+// ---- empty body falls back to summary ----
+
+func TestSinkSend_EmptyBodyFallsBackToSummary(t *testing.T) {
+	r, w := net.Pipe()
+	defer r.Close()
+	defer w.Close()
+
+	ch := make(chan frameResult, 1)
+	go readFrame(r, ch)
+
+	sess := voice.NewSessions()
+	sess.Add(w, voice.SurfacePCClient)
+	s := New(tts.NewStub(), sess)
+
+	err := s.Send(context.Background(), delivery.Sendable{
+		Body:    "",
+		Summary: "Fallback summary",
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	res := <-ch
+	if res.err != nil {
+		t.Fatalf("read frame: %v", res.err)
+	}
+
+	_, params := unmarshalFrame(t, res.buf)
+	push := unmarshalPush(t, params)
+	if push.Text != "Fallback summary" {
+		t.Errorf("text = %q, want %q", push.Text, "Fallback summary")
+	}
+}
+
+// ---- fake synths ----
+
+type errSynth struct{}
+
+func (e *errSynth) Synthesize(_ context.Context, text string) (audio.Audio, error) {
+	return audio.Audio{}, fmt.Errorf("synthesis failed")
+}
+
+func TestSinkSend_SynthesizeError(t *testing.T) {
+	sess := voice.NewSessions()
+	r, w := net.Pipe()
+	defer r.Close()
+	defer w.Close()
+	sess.Add(w, voice.SurfacePCClient)
+
+	s := New(&errSynth{}, sess)
+	err := s.Send(context.Background(), delivery.Sendable{Body: "test"})
+	if err == nil || !strings.Contains(err.Error(), "synthesize") {
+		t.Fatalf("expected synthesize error, got %v", err)
+	}
+}
+
+type badFormatSynth struct{}
+
+func (b *badFormatSynth) Synthesize(_ context.Context, text string) (audio.Audio, error) {
+	return audio.Audio{Format: audio.Format{SampleRate: 999, Channels: 0, SampleBits: 0, Encoding: "bad"}, Bytes: []byte{0, 1, 2, 3}}, nil
+}
+
+func TestSinkSend_RejectsBadAudioFormat(t *testing.T) {
+	r, w := net.Pipe()
+	defer r.Close()
+	defer w.Close()
+	sess := voice.NewSessions()
+	sess.Add(w, voice.SurfacePCClient)
+
+	s := New(&badFormatSynth{}, sess)
+	err := s.Send(context.Background(), delivery.Sendable{Body: "test"})
+	if err == nil {
+		t.Fatal("expected error for bad audio format, got nil")
+	}
+	if !strings.Contains(err.Error(), "PCM16kMono") {
+		t.Errorf("error = %q, want error containing 'PCM16kMono'", err.Error())
+	}
+}