From 388f612435f7137f361d911be640a4191de07ac6 Mon Sep 17 00:00:00 2001
From: Ariel Tov Ben <at@aisec.co.il>
Date: Tue, 21 Jul 2026 23:51:31 +0400
Subject: [PATCH] fix(discord): drain pending voice input before disconnect

---
 plugins/platforms/discord/adapter.py | 27 +++++++++++++++++
 tests/gateway/test_voice_command.py  | 43 ++++++++++++++++++++++++++++
 2 files changed, 70 insertions(+)

diff --git a/plugins/platforms/discord/adapter.py b/plugins/platforms/discord/adapter.py
index 310061439cd..e5d75d793f9 100644
--- a/plugins/platforms/discord/adapter.py
+++ b/plugins/platforms/discord/adapter.py
@@ -744,6 +744,26 @@ class VoiceReceiver:
 
         return completed
 
+    def flush_pending(self) -> list:
+        """Return buffered utterances that have not yet reached silence."""
+        completed = []
+
+        with self._lock:
+            ssrc_user_map = dict(self._ssrc_to_user)
+            for ssrc, buf in list(self._buffers.items()):
+                # 48kHz, 16-bit, stereo = 192000 bytes/sec
+                buf_duration = len(buf) / (self.SAMPLE_RATE * self.CHANNELS * 2)
+                if buf_duration >= self.MIN_SPEECH_DURATION:
+                    user_id = ssrc_user_map.get(ssrc, 0)
+                    if not user_id:
+                        user_id = self._infer_user_for_ssrc(ssrc)
+                    if user_id:
+                        completed.append((user_id, bytes(buf)))
+                self._buffers.pop(ssrc, None)
+                self._last_packet_time.pop(ssrc, None)
+
+        return completed
+
     # ------------------------------------------------------------------
     # PCM -> WAV conversion (for Whisper STT)
     # ------------------------------------------------------------------
@@ -3940,12 +3960,19 @@ class DiscordAdapter(BasePlatformAdapter):
         async with self._voice_locks.setdefault(guild_id, asyncio.Lock()):
             # Stop voice receiver first
             receiver = self._voice_receivers.pop(guild_id, None)
+            pending_inputs = []
             if receiver:
+                pending_inputs = receiver.flush_pending()
                 receiver.stop()
             listen_task = self._voice_listen_tasks.pop(guild_id, None)
             if listen_task:
                 listen_task.cancel()
 
+            guild = self._client.get_guild(guild_id) if self._client is not None else None
+            for user_id, pcm_data in pending_inputs:
+                if self._is_allowed_user(str(user_id), guild=guild, is_dm=False):
+                    await self._process_voice_input(guild_id, user_id, pcm_data)
+
             # Tear down the mixer (stops the continuous outgoing stream).
             if getattr(self, "_voice_mixers", None) is not None:
                 self._voice_mixers.pop(guild_id, None)
diff --git a/tests/gateway/test_voice_command.py b/tests/gateway/test_voice_command.py
index 6f5ca016414..8952a4d55ef 100644
--- a/tests/gateway/test_voice_command.py
+++ b/tests/gateway/test_voice_command.py
@@ -717,6 +717,18 @@ class TestVoiceReceiver:
         completed = receiver.check_silence()
         assert len(completed) == 0
 
+    def test_flush_pending_returns_recent_utterance_before_silence(self):
+        """Disconnect drains a valid utterance even before silence is detected."""
+        receiver = self._make_receiver()
+        receiver.map_ssrc(100, 42)
+        pcm_data = bytearray(b"\x00" * 96000)
+        receiver._buffers[100] = pcm_data
+        receiver._last_packet_time[100] = time.monotonic()
+
+        assert receiver.flush_pending() == [(42, bytes(pcm_data))]
+        assert 100 not in receiver._buffers
+        assert 100 not in receiver._last_packet_time
+
     def test_check_silence_unknown_user_discarded(self):
         receiver = self._make_receiver()
         # No SSRC mapping — user_id will be 0
@@ -1159,6 +1171,37 @@ class TestDiscordVoiceChannelMethods:
         assert 111 not in adapter._voice_sources
         assert 111 not in adapter._voice_receivers
 
+    @pytest.mark.asyncio
+    async def test_leave_voice_channel_processes_pending_audio_before_disconnect(self):
+        """Recent speech is transcribed before the voice connection is torn down."""
+        adapter = self._make_adapter()
+        events = []
+        mock_vc = MagicMock()
+        mock_vc.is_connected.return_value = True
+
+        async def disconnect():
+            events.append("disconnect")
+
+        mock_vc.disconnect = disconnect
+        adapter._voice_clients[111] = mock_vc
+
+        mock_receiver = MagicMock()
+        mock_receiver.flush_pending.side_effect = lambda: events.append("flush") or [(42, b"pcm")]
+        mock_receiver.stop.side_effect = lambda: events.append("stop")
+        adapter._voice_receivers[111] = mock_receiver
+        adapter._voice_listen_tasks[111] = MagicMock()
+        adapter._is_allowed_user = MagicMock(return_value=True)
+
+        async def process(guild_id, user_id, pcm_data):
+            events.append("process")
+
+        adapter._process_voice_input = process
+
+        await adapter.leave_voice_channel(111)
+
+        assert events == ["flush", "stop", "process", "disconnect"]
+        adapter._is_allowed_user.assert_called_once_with("42", guild=adapter._client.get_guild(111), is_dm=False)
+
     @pytest.mark.asyncio
     async def test_leave_voice_channel_no_connection(self):
         """Leave when not connected — no crash."""