linkgravity 1.2.2 → 1.4.0

package/src/messengers/telegram_adapter.py

@@ -0,0 +1,372 @@
+"""Telegram implementation of MessengerAdapter. 1 chat = 1 session (no
+forum-topic support yet) - see handoff notes for the forum-mode follow-up."""
+
+import asyncio
+import html
+import re
+import uuid
+from collections.abc import Awaitable, Callable
+from contextlib import asynccontextmanager, suppress
+from typing import Any
+
+from telegram import ForceReply, InlineKeyboardButton, InlineKeyboardMarkup, Message, ReactionTypeEmoji, Update
+from telegram.constants import ChatAction
+from telegram.error import TelegramError
+from telegram.ext import ContextTypes
+
+from config import logger
+from messengers.base import (
+    IncomingAttachment,
+    IncomingMessage,
+    MessengerAdapter,
+    PromptHandle,
+    ScopeOption,
+    ToolApprovalOutcome,
+)
+
+_CODE_BLOCK_RE = re.compile(r"```(?:\w+\n)?(.*?)```|`([^`\n]+)`", re.DOTALL)
+_BOLD_RE = re.compile(r"\*\*(.+?)\*\*")
+
+
+def markdown_to_telegram_html(text: str) -> str:
+    """Converts the Discord-flavored markdown this codebase generates (bold, code fences) into Telegram's HTML parse mode, escaping everything else."""
+    out = []
+    pos = 0
+    for m in _CODE_BLOCK_RE.finditer(text):
+        out.append(_BOLD_RE.sub(r"<b>\1</b>", html.escape(text[pos : m.start()])))
+        content = m.group(1) if m.group(1) is not None else m.group(2)
+        tag = "pre" if m.group(1) is not None else "code"
+        out.append(f"<{tag}>{html.escape(content)}</{tag}>")
+        pos = m.end()
+    out.append(_BOLD_RE.sub(r"<b>\1</b>", html.escape(text[pos:])))
+    return "".join(out)
+
+
+async def _safe_query_edit(query, **kwargs) -> None:
+    try:
+        await query.edit_message_text(**kwargs)
+    except TelegramError as e:
+        if "message is not modified" in str(e).lower():
+            return
+        logger.error(f"Failed to edit Telegram message via callback query: {e}")
+        raise
+
+
+class _TelegramPromptHandle(PromptHandle):
+    def __init__(self, bot, text: str, reply_markup, cleanup: Callable[[], None] | None = None):
+        self.bot = bot
+        self.text = text
+        self.reply_markup = reply_markup
+        self._cleanup = cleanup
+        self.chat_id: int | None = None
+        self.message_id: int | None = None
+        self.outcome: ToolApprovalOutcome | None = None
+
+    async def send(self, conversation_ref: int) -> Message:
+        text = self.text if len(self.text) <= 4000 else self.text[:3997] + "..."
+        try:
+            msg = await self.bot.send_message(
+                chat_id=conversation_ref, text=text, reply_markup=self.reply_markup, parse_mode="HTML"
+            )
+        except TelegramError as e:
+            logger.error(f"Failed to send Telegram prompt message: {e}")
+            raise
+        self.chat_id = msg.chat_id
+        self.message_id = msg.message_id
+        return msg
+
+    async def finalize(self) -> None:
+        if self._cleanup:
+            self._cleanup()
+        if self.message_id is None:
+            return
+        try:
+            await self.bot.edit_message_text(
+                chat_id=self.chat_id, message_id=self.message_id, text=self.text, reply_markup=None, parse_mode="HTML"
+            )
+        except TelegramError as e:
+            logger.warning(f"Failed to finalize Telegram prompt message: {e}")
+
+
+class TelegramAdapter(MessengerAdapter):
+    platform_name = "telegram"
+    supports_renaming = False
+
+    def __init__(self, bot):
+        self.bot = bot
+        # callback_data -> async handler(query); prompts add/remove their own keys here.
+        self._callbacks: dict[str, Callable[[Any], Awaitable[None]]] = {}
+
+    def register_callback(self, key: str, handler: Callable[[Any], Awaitable[None]]) -> None:
+        self._callbacks[key] = handler
+
+    async def handle_callback_query(self, update: Update, context: ContextTypes.DEFAULT_TYPE) -> None:
+        query = update.callback_query
+        if query is None or query.data is None:
+            return
+        handler = self._callbacks.pop(query.data, None)
+        if handler is None:
+            await query.answer("This action has expired.", show_alert=True)
+            return
+        await handler(query)
+
+    # -- inbound -------------------------------------------------------------
+
+    def _make_reader(self, file_id: str) -> Callable[[], Awaitable[bytes]]:
+        async def _reader() -> bytes:
+            file = await self.bot.get_file(file_id)
+            data = await file.download_as_bytearray()
+            return bytes(data)
+
+        return _reader
+
+    def to_incoming_message(self, raw_event: Update) -> IncomingMessage | None:
+        update = raw_event
+        message = update.message
+        if message is None or message.from_user is None or message.from_user.is_bot:
+            return None
+
+        attachments = []
+        if message.photo:
+            largest = message.photo[-1]
+            attachments.append(
+                IncomingAttachment(
+                    filename="photo.jpg", content_type="image/jpeg", reader=self._make_reader(largest.file_id)
+                )
+            )
+        if message.document:
+            attachments.append(
+                IncomingAttachment(
+                    filename=message.document.file_name or "file",
+                    content_type=message.document.mime_type,
+                    reader=self._make_reader(message.document.file_id),
+                )
+            )
+        if message.voice:
+            attachments.append(
+                IncomingAttachment(
+                    filename="voice.ogg",
+                    content_type=message.voice.mime_type or "audio/ogg",
+                    reader=self._make_reader(message.voice.file_id),
+                )
+            )
+        if message.audio:
+            attachments.append(
+                IncomingAttachment(
+                    filename=message.audio.file_name or "audio",
+                    content_type=message.audio.mime_type,
+                    reader=self._make_reader(message.audio.file_id),
+                )
+            )
+
+        async def add_reaction(emoji: str) -> None:
+            try:
+                await self.bot.set_message_reaction(
+                    chat_id=message.chat_id, message_id=message.message_id, reaction=[ReactionTypeEmoji(emoji=emoji)]
+                )
+            except TelegramError as e:
+                logger.warning(f"Failed to set Telegram reaction: {e}")
+
+        return IncomingMessage(
+            author_id=message.from_user.id,
+            platform=self.platform_name,
+            content=message.text or message.caption or "",
+            conversation_id=str(message.chat_id),
+            conversation_ref=message.chat_id,
+            attachments=attachments,
+            add_reaction=add_reaction,
+        )
+
+    # -- plain messaging ---------------------------------------------------
+
+    async def send_message(self, conversation_ref: int, text: str) -> Message:
+        try:
+            return await self.bot.send_message(
+                chat_id=conversation_ref, text=markdown_to_telegram_html(text), parse_mode="HTML"
+            )
+        except TelegramError as e:
+            logger.error(f"Failed to send Telegram message: {e}")
+            raise
+
+    async def edit_message(self, message_ref: Message, text: str) -> bool:
+        try:
+            await self.bot.edit_message_text(chat_id=message_ref.chat_id, message_id=message_ref.message_id, text=text)
+            return True
+        except TelegramError as e:
+            if "message is not modified" in str(e).lower():
+                return True  # already showing this exact text - not a real failure, don't send a duplicate
+            logger.warning(f"Failed to edit Telegram message: {e}")
+            return False
+
+    async def send_files(self, conversation_ref: int, file_paths: list[str]) -> None:
+        for path in file_paths:
+            with open(path, "rb") as f:
+                await self.bot.send_document(chat_id=conversation_ref, document=f)
+
+    def resolve_conversation(self, conversation_id: str) -> Any:
+        try:
+            return int(conversation_id)
+        except (TypeError, ValueError):
+            return None
+
+    async def start_conversation(self, origin_ref: int, title: str) -> int:
+        # Non-forum mode: the chat itself is the session, nothing to create.
+        return origin_ref
+
+    async def rename_conversation(self, conversation_ref: int, title: str) -> None:
+        pass  # No per-session title surface outside forum-topic mode.
+
+    @asynccontextmanager
+    async def typing(self, conversation_ref: int):
+        async def _keep_typing():
+            while True:
+                with suppress(TelegramError):
+                    await self.bot.send_chat_action(chat_id=conversation_ref, action=ChatAction.TYPING)
+                await asyncio.sleep(4)
+
+        task = asyncio.create_task(_keep_typing())
+        try:
+            yield
+        finally:
+            task.cancel()
+            with suppress(asyncio.CancelledError):
+                await task
+
+    # -- interactive prompts ----------------------------------------------
+
+    def create_tool_approval_prompt(
+        self,
+        decision_future: asyncio.Future,
+        title: str,
+        body: str,
+        scope_options: list[ScopeOption],
+    ) -> PromptHandle:
+        prompt_id = uuid.uuid4().hex[:12]
+        text = f"<b>{html.escape(title)}</b>\n\n{markdown_to_telegram_html(body)}"
+        keys: list[str] = []
+        keyboard: list[list[InlineKeyboardButton]] = []
+
+        async def resolve(decision: str, scope: ScopeOption | None, query):
+            handle.outcome = ToolApprovalOutcome(decision=decision, scope=scope)
+            if not decision_future.done():
+                decision_future.set_result(decision)
+            if decision == "allow" and scope:
+                new_text = f"✅ <b>Approved &amp; auto-allowed ({html.escape(scope.scope)})</b>"
+            elif decision == "allow":
+                new_text = "✅ <b>Approved</b>"
+            else:
+                new_text = "❌ <b>Rejected</b>"
+            handle.text = new_text
+            await query.answer()
+            await _safe_query_edit(query, text=new_text, parse_mode="HTML")
+
+        allow_key = f"{prompt_id}:allow"
+        self._callbacks[allow_key] = lambda query: resolve("allow", None, query)
+        keys.append(allow_key)
+        keyboard.append([InlineKeyboardButton("✅ Approve once", callback_data=allow_key)])
+
+        for i, opt in enumerate(scope_options):
+            suffix = " tool" if opt.kind == "tools" else ""
+            label = f"♾️ Allow [{opt.scope}]{suffix}"
+            if len(label) > 64:
+                label = label[:61] + "…"
+            key = f"{prompt_id}:scope:{i}"
+            self._callbacks[key] = lambda query, opt=opt: resolve("allow", opt, query)
+            keys.append(key)
+            keyboard.append([InlineKeyboardButton(label, callback_data=key)])
+
+        reject_key = f"{prompt_id}:reject"
+        self._callbacks[reject_key] = lambda query: resolve("reject", None, query)
+        keys.append(reject_key)
+        keyboard.append([InlineKeyboardButton("❌ Reject", callback_data=reject_key)])
+
+        handle = _TelegramPromptHandle(
+            self.bot, text, InlineKeyboardMarkup(keyboard), cleanup=lambda: [self._callbacks.pop(k, None) for k in keys]
+        )
+        return handle
+
+    def create_question_prompt(
+        self,
+        answer_future: asyncio.Future,
+        question: str,
+        options: list[str],
+        multi_select: bool = False,
+        allow_write_in: bool = True,
+    ) -> PromptHandle:
+        prompt_id = uuid.uuid4().hex[:12]
+        text = f"❓ <b>Question from AI</b>\n\n<b>{html.escape(question)}</b>\n\nPlease choose an answer below."
+        keys: list[str] = []
+
+        async def resolve(chosen_text: str, note: str, query):
+            if not answer_future.done():
+                answer_future.set_result(chosen_text)
+            new_text = f"✅ <b>{note}: {html.escape(chosen_text)}</b>"
+            handle.text = new_text
+            await query.answer()
+            await _safe_query_edit(query, text=new_text, parse_mode="HTML")
+
+        async def write_in(query):
+            await query.answer()
+            await _safe_query_edit(
+                query,
+                text=f"❓ <b>{html.escape(question)}</b>\n\n💬 Reply with your answer as a message.",
+                parse_mode="HTML",
+                reply_markup=ForceReply(selective=True),
+            )
+
+        if allow_write_in:
+            write_in_key = f"{prompt_id}:write_in"
+            self._callbacks[write_in_key] = write_in
+            keys.append(write_in_key)
+
+        if multi_select and options:
+            selected: set[int] = set()
+            shown = options[:20]
+
+            def render_keyboard() -> InlineKeyboardMarkup:
+                rows = [
+                    [
+                        InlineKeyboardButton(
+                            ("☑️ " if i in selected else "⬜ ") + opt[:60], callback_data=f"{prompt_id}:toggle:{i}"
+                        )
+                    ]
+                    for i, opt in enumerate(shown)
+                ]
+                rows.append([InlineKeyboardButton("Submit", callback_data=f"{prompt_id}:submit")])
+                if allow_write_in:
+                    rows.append([InlineKeyboardButton("✍️ Write in", callback_data=f"{prompt_id}:write_in")])
+                return InlineKeyboardMarkup(rows)
+
+            async def toggle(i: int, query):
+                selected.symmetric_difference_update({i})
+                await query.answer()
+                await query.edit_message_reply_markup(reply_markup=render_keyboard())
+
+            async def submit(query):
+                if not selected:
+                    await query.answer("Select at least one option first.", show_alert=True)
+                    return
+                chosen = ", ".join(shown[i] for i in sorted(selected))
+                await resolve(chosen, "Selected", query)
+
+            for i in range(len(shown)):
+                key = f"{prompt_id}:toggle:{i}"
+                self._callbacks[key] = lambda query, i=i: toggle(i, query)
+                keys.append(key)
+            self._callbacks[f"{prompt_id}:submit"] = submit
+            keys.append(f"{prompt_id}:submit")
+            keyboard = render_keyboard().inline_keyboard
+        else:
+            keyboard = []
+            for i, opt in enumerate(options[:20]):
+                key = f"{prompt_id}:opt:{i}"
+                self._callbacks[key] = lambda query, opt=opt: resolve(opt, "Selected", query)
+                keys.append(key)
+                keyboard.append([InlineKeyboardButton(opt[:64], callback_data=key)])
+            if allow_write_in:
+                keyboard.append([InlineKeyboardButton("✍️ Write in", callback_data=f"{prompt_id}:write_in")])
+
+        handle = _TelegramPromptHandle(
+            self.bot, text, InlineKeyboardMarkup(keyboard), cleanup=lambda: [self._callbacks.pop(k, None) for k in keys]
+        )
+        return handle

package/src/services/audio_service.py

@@ -37,7 +37,7 @@ async def tts(text: str, voice: str = None) -> bytes | None:
         os.unlink(tmp)
         return data
     except Exception as e:
-        logger.error(f"TTS error: {e}")
+        logger.exception(f"TTS error: {e}")
         return None
 
 
@@ -63,5 +63,5 @@ async def stt(audio_bytes: bytes) -> str | None:
         result = await asyncio.to_thread(_transcribe)
         return result.strip() if result else None
     except Exception as e:
-        logger.error(f"STT internal error: {e}")
+        logger.exception(f"STT internal error: {e}")
         return None

package/src/services/discord_helpers.py

@@ -3,14 +3,13 @@ import string
 import uuid
 from pathlib import Path
 
-import discord
-
 from config import TMP_FILE_DIR, logger
+from messengers.base import IncomingAttachment
 
 
-async def handle_image_attachments(message: discord.Message) -> list[str]:
+async def handle_image_attachments(attachments: list[IncomingAttachment]) -> list[str]:
     saved_paths = []
-    for att in message.attachments:
+    for att in attachments:
         ext = Path(att.filename).suffix.lower()
         filename = f"{uuid.uuid4().hex}{ext or '.tmp'}"
         dest = TMP_FILE_DIR / filename

package/src/services/response.py

@@ -2,7 +2,7 @@ from pathlib import Path
 from typing import Any
 
 from config import MAX_EMBED_LEN, MODEL_CHOICES, session_manager
-from messengers.registry import get_adapter
+from messengers.registry import get_adapter_for_platform
 from utils.utils import get_current_model
 
 
@@ -14,7 +14,7 @@ async def send_agy_response(
     start_time: float = 0,
     conv_id: str = None,
 ):
-    adapter = get_adapter()
+    adapter = get_adapter_for_platform(session.get("platform", "discord"))
     session_manager.save_sessions()
 
     parts = [response_text[i : i + MAX_EMBED_LEN] for i in range(0, max(len(response_text), 1), MAX_EMBED_LEN)]

package/src/services/streaming.py

@@ -6,25 +6,33 @@ from typing import Any
 import discord  # only for voice/TTS text cleanup below; messaging goes through the adapter
 
 from config import MAX_EMBED_LEN, STREAM_RATE_LIMIT_SEC, bot_settings, logger, session_manager
-from messengers.registry import get_adapter
+from messengers.registry import get_adapter_for_thread
 from utils.utils import clean_ansi
 
 
 def _clear_current_tool(thread_id: str):
-    """Removes the "current_tool" marker set by api/ui_routes.py while a
-    tool call is being approved - without this, the bot's presence status
-    stays stuck on the last tool after the turn finishes."""
+    """Removes the "current_tool"/"pending_approval_tool" markers set by
+    api/ui_routes.py while a tool call is being approved/run - without
+    this, the bot's presence status stays stuck on the last tool after
+    the turn finishes."""
     session = session_manager.get_session(thread_id)
-    if session and "current_tool" in session:
-        del session["current_tool"]
+    if not session:
+        return
+    changed = False
+    for key in ("current_tool", "pending_approval_tool"):
+        if key in session:
+            del session[key]
+            changed = True
+    if changed:
         session_manager.set_session(thread_id, session)
 
 
 class StreamUpdater:
-    def __init__(self, thread: Any, context_dict: dict):
+    def __init__(self, thread: Any, thread_id: str, context_dict: dict):
         self.MAX_EMBED_LEN = MAX_EMBED_LEN
         self.RATE_LIMIT_SEC = STREAM_RATE_LIMIT_SEC
         self.thread = thread
+        self.thread_id = thread_id
         self.context_dict = context_dict
         self.status_msg = None
         self.current_text = ""
@@ -57,7 +65,7 @@ class StreamUpdater:
             self.last_update_time = now
 
     async def _update(self, text: str, force_new: bool):
-        adapter = get_adapter()
+        adapter = get_adapter_for_thread(self.thread_id)
         try:
             if self.status_msg is not None and not force_new:
                 if await adapter.edit_message(self.status_msg, text):
@@ -135,8 +143,8 @@ class TTSStreamManager:
         session_manager.set_tts_task(self.thread_id, new_task)
 
 
-async def stream_thinking_latest(bot, thread: Any, context_dict: dict, queue: asyncio.Queue):
-    cog = bot.get_cog("VoiceCog")
+async def stream_thinking_latest(bot, thread: Any, thread_id: str, context_dict: dict, queue: asyncio.Queue):
+    cog = bot.get_cog("VoiceCog") if bot else None
     is_voice = bool(
         cog
         and hasattr(thread, "guild")
@@ -144,8 +152,8 @@ async def stream_thinking_latest(bot, thread: Any, context_dict: dict, queue: as
         and cog._voice_state[str(thread.guild.id)] == thread.id
     )
 
-    ui_mgr = StreamUpdater(thread, context_dict)
-    tts_mgr = TTSStreamManager(thread.id, thread.guild.id if hasattr(thread, "guild") else None, cog, is_voice)
+    ui_mgr = StreamUpdater(thread, thread_id, context_dict)
+    tts_mgr = TTSStreamManager(thread_id, thread.guild.id if hasattr(thread, "guild") else None, cog, is_voice)
 
     try:
         while True:
@@ -173,14 +181,14 @@ async def stream_thinking_latest(bot, thread: Any, context_dict: dict, queue: as
                 await ui_mgr.flush(force=True)
                 if context_dict is not None:
                     context_dict["final_text"] = ui_mgr.current_text.strip()
-                session_manager.remove_queue(str(thread.id))
-                _clear_current_tool(str(thread.id))
+                session_manager.remove_queue(thread_id)
+                _clear_current_tool(thread_id)
                 await tts_mgr.flush_all()
                 break
 
             if str(chunk).startswith("__CONV_ID__:"):
                 conv_id = chunk.split(":", 1)[1]
-                session_manager.update_session(str(thread.id), "conversation_id", conv_id)
+                session_manager.update_session(thread_id, "conversation_id", conv_id)
                 continue
 
             if chunk == "__SPLIT__":
@@ -194,6 +202,6 @@ async def stream_thinking_latest(bot, thread: Any, context_dict: dict, queue: as
             await ui_mgr.flush()
             await tts_mgr.process_chunk(chunk)
     except asyncio.CancelledError:
-        logger.debug(f"stream_thinking_latest cancelled for thread {thread.id}")
+        logger.debug(f"stream_thinking_latest cancelled for thread {thread_id}")
     except Exception as e:
         logger.exception(f"Error in stream_thinking_latest: {e}")

package/src/utils/utils.py

@@ -7,6 +7,7 @@ from core.agy_runner import (
     generate_thread_title,
     get_current_model,
     run_agy,
+    update_agy_conversation_title,
 )
 from services.audio_service import stt, tts
 from services.discord_helpers import (
@@ -23,6 +24,7 @@ __all__ = [
     "agy_send_message",
     "get_current_model",
     "generate_thread_title",
+    "update_agy_conversation_title",
     "active_processes",
     "agy_start_lock",
     "tts",

package/voice-service/index.js

@@ -271,19 +271,8 @@ function loadRustpotterModule() {
 // userId -> { rustpotter, samplesPerFrame, residual: Int16Array }
 const detectorCache = new Map();
 
-// The real pass/fail cutoff for a wake-word match. This has to be a
-// genuine, meaningful threshold, not a tuning knob to set near-zero:
-// rustpotter's confirmation logic re-arms its countdown EVERY time a
-// new "candidate" clears the threshold (detector.rs's run_detection:
-// `self.detection_countdown = self.max_mfcc_frames / 2` runs again on
-// every qualifying frame). With this set to 0.05 during earlier
-// debugging, silence and noise cleared it just as easily as real
-// speech, so the countdown never ran out and nothing was ever
-// confirmed no matter how much silence padding was fed afterward.
-// Verified against a native Rust reproduction of this exact detector
-// before settling on 0.5 (matches rustpotter's own default, and what
-// rustpotter-cli scored real captured audio at: 0.55-0.73).
-const WAKE_MATCH_THRESHOLD = 0.5;
+// Wake-word confirm cutoff - must stay well above ~0.05 (rustpotter's countdown never finalizes if noise/silence clears it too); 0.4 chosen after live use kept narrowly missing genuine hits just under 0.5.
+const WAKE_MATCH_THRESHOLD = 0.4;
 
 async function getDetectorForUser(userId) {
     if (detectorCache.has(userId)) return detectorCache.get(userId);
@@ -661,20 +650,17 @@ function setupReceiver(connection, guildId) {
                     bestDiagScoreName = diagPaddingDetection.getName();
                 }
 
-                // The real pass/fail decision - see WAKE_MATCH_THRESHOLD's
-                // comment for why this has to be a meaningful cutoff.
-                // bestDiagScore comes from the separate diagnostic
-                // instance above (see its comment in getDetectorForUser)
-                // so a "no match" line still shows the real closest
-                // score instead of a meaningless flat 0.000.
+                // Real pass/fail uses bestWakeScore; bestDiagScore is a separate, much looser detector shown only for "how close" - not on the same scale, not comparable to WAKE_MATCH_THRESHOLD.
                 wakeConfirmed = bestWakeScore >= WAKE_MATCH_THRESHOLD;
                 matchedWakeWord = wakeConfirmed ? bestWakeScoreName : null;
                 console.log(
                     wakeConfirmed
                         ? `[Wake] ${userId}: CONFIRMED (score ${bestWakeScore.toFixed(3)} for ` +
                               `"${bestWakeScoreName}", threshold ${WAKE_MATCH_THRESHOLD})`
-                        : `[Wake] ${userId}: no match (closest score ${bestDiagScore.toFixed(3)} for ` +
-                              `"${bestDiagScoreName ?? 'n/a'}", needed ${WAKE_MATCH_THRESHOLD})`,
+                        : `[Wake] ${userId}: no match (score ${bestWakeScore.toFixed(3)}, ` +
+                              `threshold ${WAKE_MATCH_THRESHOLD}; diagnostic-only closeness ` +
+                              `${bestDiagScore.toFixed(3)} for "${bestDiagScoreName ?? 'n/a'}" - ` +
+                              `different scoring config, not directly comparable to the threshold)`,
                 );
             }
 
@@ -702,7 +688,8 @@ function setupReceiver(connection, guildId) {
                 return;
             }
 
-            if (pcmBuffer.length < 24000) {
+            // Was 24000 (250ms) - cut off short Korean replies ("네"/"어"/"응"); noise is filtered upstream by isSpeaking's RMS/sustain check, not by duration.
+            if (pcmBuffer.length < 9600) {
                 if (partialSent) {
                     fetch('http://127.0.0.1:18080/stt_partial_cancel', {
                         method: 'POST',
@@ -1036,6 +1023,22 @@ process.on('uncaughtException', (err) => {
     process.exit(1);
 });
 
+// Without this, SIGTERM (lgy stop/restart) kills the process mid-connection and Discord never gets a clean leave.
+function shutdownGracefully() {
+    console.log(`[Shutdown] Disconnecting from ${connections.size} active voice connection(s)...`);
+    for (const connection of connections.values()) {
+        try {
+            connection.destroy();
+        } catch (e) {
+            // already destroyed/disconnected - fine
+        }
+    }
+    process.exit(0);
+}
+
+process.on('SIGTERM', shutdownGracefully);
+process.on('SIGINT', shutdownGracefully);
+
 const PORT = 18081;
 app.listen(PORT, '0.0.0.0', () => {
     console.log(`Node.js Voice API listening on port ${PORT}`);

package/src/services/discord_mcp.py

@@ -1,50 +0,0 @@
-import os
-
-import requests
-from mcp.server.fastmcp import FastMCP
-
-mcp = FastMCP("DiscordButtons")
-
-
-@mcp.tool()
-def ask_discord_user(question: str, options: list[str]) -> str:
-    """
-    Ask a multiple-choice question to the user in Discord using interactive buttons.
-    You MUST use this tool instead of the default ask_question tool when running in Discord.
-    """
-    thread_id = os.environ.get("DISCORD_THREAD_ID")
-    if not thread_id:
-        return "Error: DISCORD_THREAD_ID not set. Are you running in Discord?"
-
-    try:
-        resp = requests.post(
-            "http://127.0.0.1:18080/mcp_ask",
-            json={"thread_id": thread_id, "question": question, "options": options},
-            timeout=300,
-        )
-        if resp.status_code == 200:
-            return resp.json().get("answer", "No answer")
-        return f"Error: HTTP {resp.status_code}"
-    except Exception as e:
-        return f"Error: {e}"
-
-
-@mcp.tool()
-def send_discord_message(channel_id: str, message: str) -> str:
-    """
-    Sends a text message to a specific Discord channel by its ID.
-    You can use this tool when the user asks you to send a message to a different channel.
-    """
-    try:
-        resp = requests.post(
-            "http://127.0.0.1:18080/mcp_send_channel", json={"channel_id": channel_id, "message": message}, timeout=10
-        )
-        if resp.status_code == 200:
-            return "Message sent successfully"
-        return f"Error: HTTP {resp.status_code} - {resp.text}"
-    except Exception as e:
-        return f"Error: {e}"
-
-
-if __name__ == "__main__":
-    mcp.run()