fix(draft,voice): defer emoji shortcut, sequential transcription, reduce context

- Emoji-only shortcut now waits DRAFT_PROBE_DELAY before generating
  (probe text set only after delay, not while user is in chat)
- Voice transcription changed from parallel asyncio.gather to sequential
  loop to avoid Telegram FLOOD_WAIT on TranscribeAudio
- Reduce MAX_CONTEXT_MESSAGES from 40 to 30 in config.py
- Update README.md: reflect 30-message limit and sequential transcription

Author: oponfil

README.md

@@ -15,7 +15,7 @@ Auto-replies work in private chats only. Draft instructions work everywhere.
 ## Security
 
 - By default, the bot **only writes drafts** and never sends messages on your behalf. Auto-sending is only possible when the auto-reply timer is explicitly enabled in `/settings`.
-- **Messages are not stored.** The bot doesn't save conversations — chat history is fetched via Telegram API on each event and is never persisted. The context is limited to the last 40 messages or 16,000 characters (whichever comes first); older messages are dropped.
+- **Messages are not stored.** The bot doesn't save conversations — chat history is fetched via Telegram API on each event and is never persisted. The context is limited to the last 30 messages or 16,000 characters (whichever comes first); older messages are dropped.
 - **Saved Messages** (self-chat) and **Telegram service notifications** are fully ignored — the bot doesn't read, draft, or process messages in them. Additional chats can be excluded via `IGNORED_CHAT_IDS` in `config.py`.
 - Telegram sessions are encrypted with `Fernet` (`SESSION_ENCRYPTION_KEY`) before being stored in the database.
 
@@ -121,7 +121,7 @@ You can combine: `😈 tell her I miss her` — switches the style to Seducer an
 
 ### Voice Messages and Stickers
 
-All voice messages in the chat history — from both sides (yours and the contact's) — are automatically transcribed via Telegram Premium `TranscribeAudio` and included in the AI context as text. Multiple voice messages are transcribed in parallel. If transcription fails (e.g. no Premium), the message is included as `[voice message]` so the AI still knows a voice was sent. Requires Telegram Premium on the connected account (or trial attempts for free users).
+All voice messages in the chat history — from both sides (yours and the contact's) — are automatically transcribed via Telegram Premium `TranscribeAudio` and included in the AI context as text. Voice messages are transcribed sequentially to avoid Telegram rate limits; results are cached so repeated reads don't re-transcribe. If transcription fails (e.g. no Premium), the message is included as `[voice message]` so the AI still knows a voice was sent. Requires Telegram Premium on the connected account (or trial attempts for free users).
 
 Stickers are processed by emoji — the bot sees the sticker's emoji in the conversation context and generates an appropriate reply.
 

clients/pyrogram_client.py

@@ -259,16 +259,18 @@ async def read_chat_history(user_id: int, chat_id: int, limit: int = MAX_CONTEXT
                 voice_indices.append(idx)
                 voice_msg_ids.append(msg.id)
 
-        # Транскрибируем все голосовые параллельно
+        # Транскрибируем голосовые последовательно (параллельные вызовы
+        # вызывают FLOOD_WAIT от Telegram). Кэш в transcribe_voice()
+        # гарантирует, что повторные вызовы не обращаются к API.
         if voice_indices:
-            transcriptions = await asyncio.gather(
-                *(transcribe_voice(user_id, chat_id, mid) for mid in voice_msg_ids),
-            )
-            for idx, transcription in zip(voice_indices, transcriptions):
+            ok_count = 0
+            for idx, mid in zip(voice_indices, voice_msg_ids):
+                transcription = await transcribe_voice(user_id, chat_id, mid)
                 messages[idx]["text"] = transcription or "[voice message]"
+                if transcription:
+                    ok_count += 1
 
             if DEBUG_PRINT:
-                ok_count = sum(1 for t in transcriptions if t)
                 print(
                     f"{get_timestamp()} [PYROGRAM] Transcribed {ok_count}/{len(voice_indices)} "
                     f"voice messages in chat {chat_id}"

config.py

@@ -73,7 +73,7 @@
 SYSTEM_MESSAGES_FALLBACK_TTL_SECONDS = 300.0
 
 # ====== КОНТЕКСТ ======
-MAX_CONTEXT_MESSAGES = 40  # Макс. кол-во сообщений из чата для контекста
+MAX_CONTEXT_MESSAGES = 30  # Макс. кол-во сообщений из чата для контекста
 MAX_CONTEXT_CHARS = 16000  # Макс. суммарная длина текста в истории (символы)
 
 # ====== QR LOGIN ======

handlers/pyrogram_handlers.py

@@ -693,9 +693,25 @@ async def on_pyrogram_draft(user_id: int, chat_id: int, draft_text: str) -> None
         user_settings = (user or {}).get("settings") or {}
 
         if not instruction:
-            # Только emoji без инструкции — генерируем ответ как on_pyrogram_message
+            # Только emoji без инструкции — ждём выхода из чата, потом генерируем
             _cancel_auto_reply(key)
             _bot_drafts.pop(key, None)
+            _pending_drafts[key] = draft_text
+
+            await asyncio.sleep(DRAFT_PROBE_DELAY)
+
+            # Проверяем: пользователь не изменил черновик за время ожидания
+            if _pending_drafts.get(key) != draft_text:
+                return
+            _pending_drafts.pop(key, None)
+
+            if DEBUG_PRINT:
+                print(
+                    f"{get_timestamp()} [DRAFT] Processing emoji shortcut for {user_id} "
+                    f"in chat {chat_id}: {emoji_style!r}"
+                )
+
+            # Пользователь вышел — показываем пробу и генерируем
             probe_text = (await get_system_message(lang, "draft_typing")).format(emoji=STYLE_TO_EMOJI.get(emoji_style, "🦉"))
             _bot_draft_echoes[key] = probe_text
             await pyrogram_client.set_draft(user_id, chat_id, probe_text)