From ef274a482906cfd0eb1f2bf0e754b5ebd7bcc701 Mon Sep 17 00:00:00 2001
From: Alexander Russell <alex-secure@tuta.io>
Date: Thu, 16 Jul 2026 19:21:51 +1200
Subject: [PATCH] fix(tts): keep Telegram caption on the original reply text

Review follow-up: the spoken script is for synthesis only. Caption
eligibility and payload stay on the original reply, so a long reply
whose normalized script fits the 1024 char limit is still delivered
in full as its own message. Adds the long-original/short-normalized
regression case to the auto-TTS caption tests.
---
 gateway/platforms/base.py                 | 15 +++++----
 tests/gateway/test_base_topic_sessions.py | 40 +++++++++++++++++++++++
 2 files changed, 49 insertions(+), 6 deletions(-)

diff --git a/gateway/platforms/base.py b/gateway/platforms/base.py
index e816ba19f4ac..60f3344c38c8 100644
--- a/gateway/platforms/base.py
+++ b/gateway/platforms/base.py
@@ -5545,7 +5545,6 @@ class BasePlatformAdapter(ABC):
                 # an explicit ``/voice on|tts`` opt-in OR when ``voice.auto_tts`` is
                 # True globally and no ``/voice off`` has been issued.
                 _tts_path = None
-                _tts_speech_text = None
                 if (self._should_auto_tts_for_chat(event.source.chat_id)
                         and event.message_type == MessageType.VOICE
                         and text_content
@@ -5557,7 +5556,6 @@ class BasePlatformAdapter(ABC):
                             speech_text = self.prepare_tts_text(text_content)
                             if not speech_text:
                                 raise ValueError("Empty text after markdown cleanup")
-                            _tts_speech_text = speech_text
                             tts_result_str = await asyncio.to_thread(
                                 text_to_speech_tool, text=speech_text
                             )
@@ -5570,14 +5568,19 @@ class BasePlatformAdapter(ABC):
                 _tts_caption_delivered = False
                 if _tts_path and Path(_tts_path).exists():
                     try:
+                        # Caption eligibility and payload stay on the ORIGINAL
+                        # reply text. The spoken script is for synthesis only:
+                        # normalization can shrink a long reply below the
+                        # 1024-char caption limit, and captioning that spoken
+                        # form would suppress the full formatted reply the
+                        # user is meant to receive as a separate message.
                         telegram_tts_caption = None
-                        caption_text = _tts_speech_text or self.prepare_tts_text(text_content)
                         if (
                             self.platform == Platform.TELEGRAM
-                            and caption_text
-                            and caption_text[:1024] == caption_text
+                            and text_content
+                            and text_content[:1024] == text_content
                         ):
-                            telegram_tts_caption = caption_text
+                            telegram_tts_caption = text_content
                         tts_result = await self.play_tts(
                             chat_id=event.source.chat_id,
                             audio_path=_tts_path,
diff --git a/tests/gateway/test_base_topic_sessions.py b/tests/gateway/test_base_topic_sessions.py
index 03934d69d629..e7ce6c1f52b2 100644
--- a/tests/gateway/test_base_topic_sessions.py
+++ b/tests/gateway/test_base_topic_sessions.py
@@ -333,6 +333,46 @@ class TestTelegramAutoTtsCaptionDelivery:
             }
         ]
 
+    @pytest.mark.asyncio
+    async def test_long_original_with_short_spoken_script_still_sends_full_reply(self, tmp_path):
+        adapter = DummyTelegramAdapter()
+        adapter._keep_typing = self._hold_typing()
+        adapter._should_auto_tts_for_chat = lambda _chat_id: True
+        adapter.play_tts = AsyncMock(return_value=SendResult(success=True, message_id="tts-1"))
+        # Markdown-heavy reply: over the 1024-char caption limit as written,
+        # but the normalized spoken script (markdown and URLs removed) is far
+        # below it. Caption eligibility must follow the ORIGINAL reply, so the
+        # full formatted text is still delivered as its own message instead of
+        # being swallowed into a lossy caption.
+        long_reply = "\n".join(
+            f"- **item {i}** [details](https://example.com/some/very/long/path/{i:04d})"
+            for i in range(20)
+        )
+        assert len(long_reply) > 1024
+        assert len(adapter.prepare_tts_text(long_reply)) <= 1024
+        adapter.set_message_handler(lambda _event: asyncio.sleep(0, result=long_reply))
+
+        tts_path = tmp_path / "reply.ogg"
+        tts_path.write_text("audio", encoding="utf-8")
+        event = self._make_voice_event()
+
+        with patch("tools.tts_tool.check_tts_requirements", return_value=True), patch(
+            "tools.tts_tool.text_to_speech_tool",
+            return_value=json.dumps({"file_path": str(tts_path)}),
+        ):
+            await adapter._process_message_background(event, build_session_key(event.source))
+
+        adapter.play_tts.assert_awaited_once()
+        assert adapter.play_tts.await_args.kwargs["caption"] is None
+        assert adapter.sent == [
+            {
+                "chat_id": "-1001",
+                "content": long_reply,
+                "reply_to": None,
+                "metadata": {"thread_id": "17585", "notify": True},
+            }
+        ]
+
     @pytest.mark.asyncio
     async def test_telegram_auto_tts_send_failure_keeps_followup_text(self, tmp_path):
         adapter = DummyTelegramAdapter()