feat(voice): rework TTS/ASR stack and unify tool/skill config schema

agent/memory/conversation_store.py

@@ -44,6 +44,7 @@ CREATE TABLE IF NOT EXISTS messages (
     role         TEXT    NOT NULL,
     content      TEXT    NOT NULL,
     created_at   INTEGER NOT NULL,
+    extras       TEXT    NOT NULL DEFAULT '',
     UNIQUE (session_id, seq)
 );
 
@@ -67,6 +68,12 @@ _MIGRATION_ADD_CONTEXT_START_SEQ = """
 ALTER TABLE sessions ADD COLUMN context_start_seq INTEGER NOT NULL DEFAULT 0;
 """
 
+# Generic JSON sidecar for per-message attachments (TTS audio URL, future use).
+# Always optional — readers must tolerate missing column / empty / invalid JSON.
+_MIGRATION_ADD_MSG_EXTRAS = """
+ALTER TABLE messages ADD COLUMN extras TEXT NOT NULL DEFAULT '';
+"""
+
 DEFAULT_MAX_AGE_DAYS: int = 30
 
 
@@ -169,20 +176,26 @@ def _group_into_display_turns(
     cur_rest: List[tuple] = []
     started = False
 
-    for role, raw_content, created_at in rows:
+    for role, raw_content, created_at, raw_extras in rows:
         try:
             content = json.loads(raw_content)
         except Exception:
             content = raw_content
+        try:
+            extras = json.loads(raw_extras) if raw_extras else {}
+            if not isinstance(extras, dict):
+                extras = {}
+        except Exception:
+            extras = {}
 
         if role == "user" and _is_visible_user_message(content):
             if started:
                 groups.append((cur_user, cur_rest))
-            cur_user = (content, created_at)
+            cur_user = (content, created_at, extras)
             cur_rest = []
             started = True
         else:
-            cur_rest.append((role, content, created_at))
+            cur_rest.append((role, content, created_at, extras))
 
     if started:
         groups.append((cur_user, cur_rest))
@@ -195,7 +208,7 @@ def _group_into_display_turns(
     for user_row, rest in groups:
         # User turn
         if user_row:
-            content, created_at = user_row
+            content, created_at, _u_extras = user_row
             text = _extract_display_text(content)
             if text:
                 turns.append({"role": "user", "content": text, "created_at": created_at})
@@ -206,8 +219,11 @@ def _group_into_display_turns(
         tool_results: Dict[str, str] = {}
         final_text = ""
         final_ts: Optional[int] = None
+        merged_extras: Dict[str, Any] = {}
 
-        for role, content, created_at in rest:
+        for role, content, created_at, extras in rest:
+            if role == "assistant" and isinstance(extras, dict):
+                merged_extras.update(extras)
             if role == "user":
                 tool_results.update(_extract_tool_results(content))
             elif role == "assistant":
@@ -256,6 +272,8 @@ def _group_into_display_turns(
                 "steps": steps,
                 "created_at": final_ts or (user_row[1] if user_row else 0),
             }
+            if merged_extras:
+                turn["extras"] = merged_extras
             turns.append(turn)
 
     return turns
@@ -411,13 +429,15 @@ class ConversationStore:
                         content = json.dumps(
                             msg.get("content", ""), ensure_ascii=False
                         )
+                        extras_obj = msg.get("extras") or {}
+                        extras = json.dumps(extras_obj, ensure_ascii=False) if extras_obj else ""
                         conn.execute(
                             """
                             INSERT OR IGNORE INTO messages
-                                (session_id, seq, role, content, created_at)
-                            VALUES (?, ?, ?, ?, ?)
+                                (session_id, seq, role, content, created_at, extras)
+                            VALUES (?, ?, ?, ?, ?, ?)
                             """,
-                            (session_id, next_seq, role, content, now),
+                            (session_id, next_seq, role, content, now, extras),
                         )
                         next_seq += 1
 
@@ -651,6 +671,55 @@ class ConversationStore:
             logger.info(f"[ConversationStore] Pruned {deleted} expired sessions")
         return deleted
 
+    def attach_extras_to_last_assistant(
+        self,
+        session_id: str,
+        extras: Dict[str, Any],
+    ) -> Optional[int]:
+        """
+        Merge ``extras`` into the latest assistant message of a session.
+
+        Used by post-processing (e.g. TTS) that needs to annotate an already
+        persisted bot reply with attachments such as audio URLs.
+
+        Returns the message seq that was updated, or ``None`` if no assistant
+        message exists or the update could not be applied.
+        """
+        if not extras:
+            return None
+        with self._lock:
+            conn = self._connect()
+            try:
+                row = conn.execute(
+                    """
+                    SELECT seq, extras FROM messages
+                    WHERE session_id = ? AND role = 'assistant'
+                    ORDER BY seq DESC LIMIT 1
+                    """,
+                    (session_id,),
+                ).fetchone()
+                if not row:
+                    return None
+                seq, raw = row
+                try:
+                    cur = json.loads(raw) if raw else {}
+                    if not isinstance(cur, dict):
+                        cur = {}
+                except Exception:
+                    cur = {}
+                cur.update(extras)
+                conn.execute(
+                    "UPDATE messages SET extras = ? WHERE session_id = ? AND seq = ?",
+                    (json.dumps(cur, ensure_ascii=False), session_id, seq),
+                )
+                conn.commit()
+                return seq
+            except Exception as e:
+                logger.warning(f"[ConversationStore] attach_extras failed: {e}")
+                return None
+            finally:
+                conn.close()
+
     def load_history_page(
         self,
         session_id: str,
@@ -698,7 +767,22 @@ class ConversationStore:
                 ).fetchone()
                 ctx_start = ctx_row[0] if ctx_row else 0
 
+                # extras column is added by migration; tolerate older DBs that
+                # might miss it by falling back to a NULL literal.
+                try:
                     rows = conn.execute(
+                        """
+                        SELECT seq, role, content, created_at, extras
+                        FROM messages
+                        WHERE session_id = ?
+                        ORDER BY seq ASC
+                        """,
+                        (session_id,),
+                    ).fetchall()
+                except sqlite3.OperationalError:
+                    rows = [
+                        (seq, role, content, created_at, "")
+                        for (seq, role, content, created_at) in conn.execute(
                             """
                             SELECT seq, role, content, created_at
                             FROM messages
@@ -707,6 +791,7 @@ class ConversationStore:
                             """,
                             (session_id,),
                         ).fetchall()
+                    ]
             finally:
                 conn.close()
 
@@ -719,13 +804,16 @@ class ConversationStore:
             include_thinking = False
 
         # Strip seq for display grouping, but record max seq per visible user group
-        plain_rows = [(role, content, created_at) for _seq, role, content, created_at in rows]
+        plain_rows = [
+            (role, content, created_at, extras_raw)
+            for _seq, role, content, created_at, extras_raw in rows
+        ]
         visible = _group_into_display_turns(plain_rows, include_thinking=include_thinking)
 
         # Build a mapping: find the seq of each visible user message to annotate context boundary.
         # Walk through rows to find visible user message seqs in order.
         visible_user_seqs: List[int] = []
-        for seq, role, raw_content, _ts in rows:
+        for seq, role, raw_content, _ts, _extras in rows:
             if role != "user":
                 continue
             try:
@@ -911,6 +999,18 @@ class ConversationStore:
             except Exception as e:
                 logger.warning(f"[ConversationStore] Migration (context_start_seq) failed: {e}")
 
+        msg_cols = {
+            row[1]
+            for row in conn.execute("PRAGMA table_info(messages)").fetchall()
+        }
+        if "extras" not in msg_cols:
+            try:
+                conn.execute(_MIGRATION_ADD_MSG_EXTRAS)
+                conn.commit()
+                logger.info("[ConversationStore] Migrated: added messages.extras column")
+            except Exception as e:
+                logger.warning(f"[ConversationStore] Migration (extras) failed: {e}")
+
     def _connect(self) -> sqlite3.Connection:
         conn = sqlite3.connect(str(self._db_path), timeout=10)
         conn.execute("PRAGMA journal_mode=WAL")

agent/tools/vision/vision.py

@@ -3,7 +3,7 @@ Vision tool - Analyze images using Vision API.
 Supports local files (auto base64-encoded) and HTTP URLs.
 
 Provider resolution:
-  - tool.vision.model (if set) means "prefer this model first; fall back to
+  - tools.vision.model (if set) means "prefer this model first; fall back to
     other configured providers if it fails". The model name is mapped to its
     native provider (e.g. doubao-* → Doubao, kimi-* → Moonshot, gpt-* →
     OpenAI/LinkAI). That provider is tried first, then the standard auto
@@ -60,7 +60,7 @@ _DISCOVERABLE_MODELS = [
 ]
 
 # Model name prefix → discoverable provider display_name.
-# Used to auto-route tool.vision.model to its native provider.
+# Used to auto-route tools.vision.model to its native provider.
 # Matched case-insensitively; longest prefix wins.
 _MODEL_PREFIX_TO_PROVIDER = [
     ("doubao-", "Doubao"),
@@ -154,7 +154,7 @@ class Vision(BaseTool):
 
         # Default model is only used as a last-resort placeholder for providers
         # whose VisionProvider.model_override is None (e.g. raw OpenAI provider
-        # when the user did not configure tool.vision.model).
+        # when the user did not configure tools.vision.model).
         return self._call_with_fallback(providers, DEFAULT_MODEL, question, image_content)
 
     def _call_with_fallback(self, providers: List[VisionProvider], model: str,
@@ -193,12 +193,12 @@ class Vision(BaseTool):
         """
         Build an ordered list of providers to try.
 
-        Semantics of `tool.vision.model`:
+        Semantics of `tools.vision.model`:
           "Prefer this model first; fall back to other configured providers
            if it fails."
 
         Order:
-          1. The provider that natively serves `tool.vision.model` (if any
+          1. The provider that natively serves `tools.vision.model` (if any
              and its API key is configured) — using the user-specified model
              name verbatim.
           2. Auto-discovery chain as fallback:
@@ -213,7 +213,7 @@ class Vision(BaseTool):
         user_model = self._resolve_user_vision_model()
         providers: List[VisionProvider] = []
 
-        # Step 1: preferred provider derived from tool.vision.model
+        # Step 1: preferred provider derived from tools.vision.model
         if user_model:
             preferred = self._route_by_model_name(user_model)
             if preferred:
@@ -251,11 +251,11 @@ class Vision(BaseTool):
 
     @staticmethod
     def _resolve_user_vision_model() -> Optional[str]:
-        """Read tool.vision.model from config; return None if unset/blank."""
-        tool_conf = conf().get("tool", {})
-        if not isinstance(tool_conf, dict):
+        """Read tools.vision.model (singular ``tool`` kept as runtime fallback)."""
+        tools_conf = conf().get("tools") or conf().get("tool") or {}
+        if not isinstance(tools_conf, dict):
             return None
-        vision_conf = tool_conf.get("vision", {})
+        vision_conf = tools_conf.get("vision", {})
         if not isinstance(vision_conf, dict):
             return None
         m = vision_conf.get("model")
@@ -303,7 +303,7 @@ class Vision(BaseTool):
                 self._append_provider(providers, lambda: self._build_linkai_provider(user_model))
             if providers:
                 return providers
-            logger.warning(f"[Vision] tool.vision.model='{user_model}' looks like an OpenAI "
+            logger.warning(f"[Vision] tools.vision.model='{user_model}' looks like an OpenAI "
                            f"model but neither OPENAI_API_KEY nor LINKAI_API_KEY is configured.")
             return None  # fall through to auto
 
@@ -317,7 +317,7 @@ class Vision(BaseTool):
                 continue
             api_key = conf().get(config_key, "")
             if not api_key or not api_key.strip():
-                logger.warning(f"[Vision] tool.vision.model='{user_model}' routes to "
+                logger.warning(f"[Vision] tools.vision.model='{user_model}' routes to "
                                f"'{display_name}' but '{config_key}' is not configured. "
                                f"Falling back to auto-discovery.")
                 return None  # fall through to auto
@@ -452,8 +452,8 @@ class Vision(BaseTool):
         if not self._main_bot_supports_vision(bot):
             return None
 
-        # Use the configured main model name; do NOT inject tool.vision.model
-        # here, because by the time we reach this branch the tool.vision.model
+        # Use the configured main model name; do NOT inject tools.vision.model
+        # here, because by the time we reach this branch the tools.vision.model
         # routing has already been attempted (and either matched the main bot
         # or failed to find a provider).
         main_model_name = conf().get("model") or None

channel/chat_channel.py

@@ -171,7 +171,13 @@ class ChatChannel(Channel):
             if "desire_rtype" not in context and conf().get("always_reply_voice") and ReplyType.VOICE not in self.NOT_SUPPORT_REPLYTYPE:
                 context["desire_rtype"] = ReplyType.VOICE
         elif context.type == ContextType.VOICE:
-            if "desire_rtype" not in context and conf().get("voice_reply_voice") and ReplyType.VOICE not in self.NOT_SUPPORT_REPLYTYPE:
+            # Voice input replies with voice when either voice_reply_voice
+            # (mirror voice) or the global always_reply_voice toggle is on.
+            if (
+                "desire_rtype" not in context
+                and (conf().get("voice_reply_voice") or conf().get("always_reply_voice"))
+                and ReplyType.VOICE not in self.NOT_SUPPORT_REPLYTYPE
+            ):
                 context["desire_rtype"] = ReplyType.VOICE
         return context
 

channel/feishu/feishu_channel.py

@@ -1515,10 +1515,16 @@ class FeiShuChanel(ChatChannel):
             else:
                 context.type = ContextType.TEXT
             context.content = content.strip()
+            # Text input opts into voice replies only when the always-on toggle is set.
+            if "desire_rtype" not in context and conf().get("always_reply_voice"):
+                context["desire_rtype"] = ReplyType.VOICE
 
         elif context.type == ContextType.VOICE:
-            # 2.语音请求
-            if "desire_rtype" not in context and conf().get("voice_reply_voice"):
+            # 2.语音请求: voice input replies with voice if either
+            # voice_reply_voice (mirror reply) or always_reply_voice is on.
+            if "desire_rtype" not in context and (
+                conf().get("voice_reply_voice") or conf().get("always_reply_voice")
+            ):
                 context["desire_rtype"] = ReplyType.VOICE
 
         return context

channel/web/static/css/console.css

@@ -1294,3 +1294,76 @@
     overflow: hidden;
     min-height: 2.5em;  /* ~2 lines at text-sm leading-relaxed */
 }
+
+/* --------------------------------------------------------------------
+ * Voice pill — compact custom audio player used by mic uploads and TTS
+ * replies. Replaces the bulky native <audio controls> with a play/pause
+ * icon + thin progress bar + duration counter so it blends into chat
+ * bubbles without the chrome-grey browser default look.
+ * ------------------------------------------------------------------ */
+.voice-pill {
+    display: inline-flex;
+    align-items: center;
+    gap: 8px;
+    padding: 6px 10px;
+    border-radius: 999px;
+    background: rgba(15, 23, 42, 0.05);
+    color: rgb(71, 85, 105);
+    font-size: 12px;
+    line-height: 1;
+    max-width: 240px;
+    user-select: none;
+    cursor: default;
+}
+.dark .voice-pill {
+    background: rgba(255, 255, 255, 0.08);
+    color: rgb(203, 213, 225);
+}
+.voice-pill[data-loading="1"] {
+    opacity: 0.65;
+}
+.voice-pill-btn {
+    width: 22px;
+    height: 22px;
+    border-radius: 999px;
+    display: inline-flex;
+    align-items: center;
+    justify-content: center;
+    background: var(--color-primary-500, #2563eb);
+    color: #fff;
+    flex-shrink: 0;
+    cursor: pointer;
+    transition: transform 0.1s ease;
+}
+.voice-pill-btn:hover { transform: scale(1.05); }
+.voice-pill-btn i { font-size: 9px; margin-left: 1px; }
+.voice-pill-btn[data-state="play"] i { margin-left: 2px; }
+.voice-pill-btn[data-state="pause"] i { margin-left: 0; }
+.voice-pill-track {
+    flex: 1;
+    height: 3px;
+    border-radius: 999px;
+    background: rgba(100, 116, 139, 0.25);
+    overflow: hidden;
+    min-width: 70px;
+}
+.dark .voice-pill-track {
+    background: rgba(148, 163, 184, 0.25);
+}
+.voice-pill-fill {
+    height: 100%;
+    width: 0%;
+    background: var(--color-primary-500, #2563eb);
+    border-radius: inherit;
+    transition: width 0.1s linear;
+}
+.voice-pill-time {
+    font-variant-numeric: tabular-nums;
+    font-size: 11px;
+    color: inherit;
+    opacity: 0.75;
+    flex-shrink: 0;
+    min-width: 28px;
+    text-align: right;
+}
+.voice-pill audio { display: none; }

channel/web/static/js/console.js

@@ -25,6 +25,7 @@ const I18N = {
         models_add_vendor: '添加厂商',
         models_provider: '厂商',
         models_model: '模型',
+        models_voice: '音色',
         models_configured: '已配置',
         models_not_configured: '未配置',
         models_pick_to_configure: '选择以配置',
@@ -160,6 +161,11 @@ const I18N = {
         mic_permission_denied: '无法访问麦克风，请检查浏览器权限',
         mic_too_short: '录音太短，请重试',
         mic_error: '语音识别失败',
+        speak_msg: '朗读这段回复',
+        voice_reply_mode_label: '语音回复策略',
+        voice_reply_off: '关闭',
+        voice_reply_if_voice: '仅语音问/语音答',
+        voice_reply_always: '总是语音回复',
         attach_menu_folder: '上传文件夹',
         confirm_yes: '确认',
         confirm_cancel: '取消',
@@ -180,6 +186,7 @@ const I18N = {
         models_add_vendor: 'Add Vendor',
         models_provider: 'Provider',
         models_model: 'Model',
+        models_voice: 'Voice',
         models_configured: 'configured',
         models_not_configured: 'not configured',
         models_pick_to_configure: 'pick to configure',
@@ -315,6 +322,11 @@ const I18N = {
         mic_permission_denied: 'Cannot access microphone — check browser permissions',
         mic_too_short: 'Recording too short, please retry',
         mic_error: 'Speech recognition failed',
+        speak_msg: 'Read this reply aloud',
+        voice_reply_mode_label: 'Voice reply policy',
+        voice_reply_off: 'Off',
+        voice_reply_if_voice: 'Voice only if voice input',
+        voice_reply_always: 'Always reply with voice',
         attach_menu_folder: 'Upload Folder',
         confirm_yes: 'Confirm',
         confirm_cancel: 'Cancel',
@@ -1474,6 +1486,7 @@ function sendVoiceMessage(text, audioUrl) {
         message: text,
         stream: true,
         timestamp: timestamp.toISOString(),
+        is_voice: true,
     };
 
     const MAX_RETRIES = 2;
@@ -1512,19 +1525,19 @@ function sendVoiceMessage(text, audioUrl) {
 function addUserVoiceMessage(audioUrl, caption, timestamp) {
     const el = document.createElement('div');
     el.className = 'flex justify-end px-4 sm:px-6 py-3';
-    // Voice-message bubble: playable <audio> on top, ASR caption beneath.
+    // Voice-message bubble: compact voice pill on top, ASR caption beneath.
     // The bubble keeps the same primary tint as a normal user message so
     // it visually slots into the conversation flow.
     el.innerHTML = `
         <div class="max-w-[75%] sm:max-w-[60%]">
             <div class="bg-slate-100 dark:bg-white/10 text-slate-700 dark:text-slate-200 rounded-2xl px-3 py-2 msg-content user-bubble">
-                <audio controls preload="metadata" src="${audioUrl}"
-                       class="block w-[260px] max-w-full h-9"></audio>
+                <div class="user-voice-slot"></div>
                 ${caption ? `<div class="text-xs mt-1.5 leading-snug text-slate-500 dark:text-slate-400 whitespace-pre-wrap break-words">${escapeHtml(caption)}</div>` : ''}
             </div>
             <div class="text-xs text-slate-400 dark:text-slate-500 mt-1.5 text-right">${formatTime(timestamp)}</div>
         </div>
     `;
+    el.querySelector('.user-voice-slot').appendChild(renderVoicePill(audioUrl));
     messagesDiv.appendChild(el);
     _autoScrollEnabled = true;
     scrollChatToBottom(true);
@@ -1639,12 +1652,16 @@ function startSSE(requestId, loadingEl, timestamp, titleInfo) {
                     <div class="agent-steps"></div>
                     <div class="answer-content sse-streaming"></div>
                     <div class="media-content"></div>
+                    <div class="bot-audio-slot"></div>
                 </div>
                 <div class="flex items-center gap-2 mt-1.5">
                     <span class="text-xs text-slate-400 dark:text-slate-500">${formatTime(timestamp)}</span>
                     <button class="copy-msg-btn text-xs text-slate-300 dark:text-slate-600 hover:text-slate-500 dark:hover:text-slate-400 transition-colors cursor-pointer" title="${currentLang === 'zh' ? '复制' : 'Copy'}" style="display:none">
                         <i class="fas fa-copy"></i>
                     </button>
+                    <button class="speak-msg-btn text-xs text-slate-300 dark:text-slate-600 hover:text-slate-500 dark:hover:text-slate-400 transition-colors cursor-pointer" title="${t('speak_msg')}" style="display:none;">
+                        <i class="fas fa-volume-up"></i>
+                    </button>
                 </div>
             </div>
         `;
@@ -1856,11 +1873,12 @@ function startSSE(requestId, loadingEl, timestamp, titleInfo) {
                 scrollChatToBottom();
 
             } else if (item.type === 'done') {
+                // Don't close the stream yet: the backend keeps it open
+                // for a short tail to deliver async attachments such as
+                // TTS audio (`voice_attach`). It will close the stream on
+                // its own via onerror once the tail expires.
                 done = true;
-                es.close();
-                delete activeStreams[requestId];
 
-                // item.content may be empty when "done" is only a stream-close signal after media.
                 const finalText = item.content || accumulatedText;
 
                 if (!botEl && finalText) {
@@ -1874,6 +1892,7 @@ function startSSE(requestId, loadingEl, timestamp, titleInfo) {
                     if (copyBtn && finalText) copyBtn.style.display = '';
                     applyHighlighting(botEl);
                 }
+                renderBotSpeakerButton(botEl, finalText);
                 scrollChatToBottom();
 
                 if (titleInfo) {
@@ -1883,6 +1902,15 @@ function startSSE(requestId, loadingEl, timestamp, titleInfo) {
                     loadSessionList();
                 }
 
+            } else if (item.type === 'voice_attach') {
+                // TTS finished — attach a playable audio element to the
+                // current bot bubble. The stream closes right after.
+                if (botEl && item.url) {
+                    attachAudioToBotBubble(botEl, item.url, { autoplay: true });
+                }
+                es.close();
+                delete activeStreams[requestId];
+
             } else if (item.type === 'error') {
                 done = true;
                 es.close();
@@ -1896,7 +1924,10 @@ function startSSE(requestId, loadingEl, timestamp, titleInfo) {
             es.close();
             delete activeStreams[requestId];
 
-            if (done) return;
+            if (done) {
+                // Normal close after the post-done tail expired; nothing to do.
+                return;
+            }
 
             if (currentReasoningEl) {
                 finalizeThinking(currentReasoningEl, reasoningStartTime, reasoningText);
@@ -2187,21 +2218,174 @@ function createBotMessageEl(content, timestamp, requestId, msg) {
             <div class="bg-white dark:bg-[#1A1A1A] border border-slate-200 dark:border-white/10 rounded-2xl px-4 py-3 text-sm leading-relaxed msg-content text-slate-700 dark:text-slate-200">
                 ${stepsHtml ? `<div class="agent-steps">${stepsHtml}</div>` : ''}
                 <div class="answer-content">${renderMarkdown(displayContent)}</div>
+                <div class="bot-audio-slot"></div>
             </div>
             <div class="flex items-center gap-2 mt-1.5">
                 <span class="text-xs text-slate-400 dark:text-slate-500">${formatTime(timestamp)}</span>
                 <button class="copy-msg-btn text-xs text-slate-300 dark:text-slate-600 hover:text-slate-500 dark:hover:text-slate-400 transition-colors cursor-pointer" title="${currentLang === 'zh' ? '复制' : 'Copy'}">
                     <i class="fas fa-copy"></i>
                 </button>
+                <button class="speak-msg-btn text-xs text-slate-300 dark:text-slate-600 hover:text-slate-500 dark:hover:text-slate-400 transition-colors cursor-pointer" title="${t('speak_msg')}" style="display:none;">
+                    <i class="fas fa-volume-up"></i>
+                </button>
             </div>
         </div>
     `;
     el.querySelector('.answer-content').dataset.rawMd = displayContent;
+    // Existing TTS attachment (history replay): mount the player up-front.
+    const existingAudio = msg && msg.extras && msg.extras.audio && msg.extras.audio.url;
+    if (existingAudio) {
+        attachAudioToBotBubble(el, existingAudio, { autoplay: false });
+    }
+    renderBotSpeakerButton(el, displayContent);
     applyHighlighting(el);
     bindChatKnowledgeLinks(el);
     return el;
 }
 
+// Append (or replace) a small audio player inside a bot bubble's
+// dedicated `.bot-audio-slot`. Used by both live TTS pushes and history
+// replay. Silent failures: never throws.
+function attachAudioToBotBubble(botEl, audioUrl, opts) {
+    try {
+        if (!botEl || !audioUrl) return;
+        const slot = botEl.querySelector('.bot-audio-slot');
+        if (!slot) return;
+        slot.innerHTML = '';
+        slot.style.marginTop = '6px';
+        const pill = renderVoicePill(audioUrl, { autoplay: !!(opts && opts.autoplay) });
+        slot.appendChild(pill);
+        const speakBtn = botEl.querySelector('.speak-msg-btn');
+        if (speakBtn) speakBtn.style.display = 'none';
+    } catch (_) { /* silent */ }
+}
+
+// Build a compact play/pause + progress + duration pill that wraps a
+// hidden <audio>. Returns the root element; safe to embed anywhere.
+function renderVoicePill(audioUrl, opts) {
+    opts = opts || {};
+    const wrap = document.createElement('div');
+    wrap.className = 'voice-pill';
+    wrap.innerHTML = `
+        <button type="button" class="voice-pill-btn" data-state="play" aria-label="play">
+            <i class="fas fa-play"></i>
+        </button>
+        <div class="voice-pill-track"><div class="voice-pill-fill"></div></div>
+        <span class="voice-pill-time">0:00</span>
+        <audio preload="metadata" src="${audioUrl}"></audio>
+    `;
+    const btn = wrap.querySelector('.voice-pill-btn');
+    const fill = wrap.querySelector('.voice-pill-fill');
+    const timeEl = wrap.querySelector('.voice-pill-time');
+    const audio = wrap.querySelector('audio');
+
+    const fmt = (s) => {
+        if (!isFinite(s) || s < 0) s = 0;
+        const m = Math.floor(s / 60);
+        const r = Math.floor(s % 60);
+        return `${m}:${r < 10 ? '0' : ''}${r}`;
+    };
+    const setIcon = (state) => {
+        btn.dataset.state = state;
+        btn.querySelector('i').className = state === 'pause' ? 'fas fa-pause' : 'fas fa-play';
+        btn.setAttribute('aria-label', state === 'pause' ? 'pause' : 'play');
+    };
+
+    audio.addEventListener('loadedmetadata', () => {
+        if (audio.duration && isFinite(audio.duration)) timeEl.textContent = fmt(audio.duration);
+    });
+    audio.addEventListener('timeupdate', () => {
+        const dur = audio.duration || 0;
+        if (dur > 0) {
+            fill.style.width = `${Math.min(100, (audio.currentTime / dur) * 100)}%`;
+            timeEl.textContent = fmt(dur - audio.currentTime);
+        }
+    });
+    audio.addEventListener('ended', () => {
+        setIcon('play');
+        fill.style.width = '0%';
+        timeEl.textContent = fmt(audio.duration || 0);
+    });
+    audio.addEventListener('play',  () => setIcon('pause'));
+    audio.addEventListener('pause', () => setIcon('play'));
+
+    btn.addEventListener('click', (e) => {
+        e.stopPropagation();
+        if (audio.paused) {
+            audio.play().catch(() => {});
+        } else {
+            audio.pause();
+        }
+    });
+
+    if (opts.autoplay) {
+        // Autoplay may be blocked by the browser; fall back silently and
+        // let the user tap the play button.
+        const tryPlay = () => audio.play().catch(() => {});
+        if (audio.readyState >= 2) tryPlay();
+        else audio.addEventListener('canplay', tryPlay, { once: true });
+    }
+    return wrap;
+}
+
+// Show the manual "read aloud" button when TTS is configured but the
+// bubble has no audio yet. Lazily probes capability via /api/models so
+// we don't expose the button when nothing can synthesize speech.
+function renderBotSpeakerButton(botEl, text) {
+    if (!botEl || !text || !text.trim()) return;
+    const btn = botEl.querySelector('.speak-msg-btn');
+    if (!btn) return;
+    if (botEl.querySelector('.bot-audio-slot audio')) return;
+    _isTtsReady().then(ready => {
+        if (!ready) return;
+        btn.style.display = '';
+        btn.onclick = () => _triggerManualTts(btn, botEl, text);
+    });
+}
+
+let _ttsReadyPromise = null;
+let _ttsReadyTs = 0;
+function _isTtsReady() {
+    // Cache for 30s to avoid hammering /api/models on every bubble.
+    if (_ttsReadyPromise && Date.now() - _ttsReadyTs < 30000) {
+        return _ttsReadyPromise;
+    }
+    _ttsReadyTs = Date.now();
+    _ttsReadyPromise = fetch('/api/models')
+        .then(r => r.json())
+        .then(data => {
+            const tts = data && data.capabilities && data.capabilities.tts;
+            if (!tts) return false;
+            return Boolean(tts.current_provider || tts.suggested_provider);
+        })
+        .catch(() => false);
+    return _ttsReadyPromise;
+}
+
+function _triggerManualTts(btn, botEl, text) {
+    if (btn.dataset.busy === '1') return;
+    btn.dataset.busy = '1';
+    const icon = btn.querySelector('i');
+    const prev = icon ? icon.className : '';
+    if (icon) icon.className = 'fas fa-spinner fa-spin';
+    fetch('/api/voice/tts', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ text, session_id: sessionId }),
+    })
+        .then(r => r.json())
+        .then(data => {
+            if (data && data.status === 'success' && data.audio_url) {
+                attachAudioToBotBubble(botEl, data.audio_url, { autoplay: true });
+            }
+        })
+        .catch(() => {})
+        .finally(() => {
+            btn.dataset.busy = '0';
+            if (icon) icon.className = prev || 'fas fa-volume-up';
+        });
+}
+
 function addUserMessage(content, timestamp, attachments) {
     const el = createUserMessageEl(content, timestamp, attachments);
     messagesDiv.appendChild(el);
@@ -3842,14 +4026,39 @@ function renderCapabilityBody(def, cap, body) {
 
     body.innerHTML = providerHtml + modelHtml + dimHtml + footer;
 
-    // The body subtree is detached from `document` at this moment (the parent
-    // wrap is not yet appended), so we must scope lookups to `body` rather
-    // than calling document.getElementById, which would return null and crash
-    // initDropdown's internal querySelector.
+    // TTS: mount reply-mode above provider; defer off-mode toggle to the end.
+    if (def.id === 'tts') {
+        renderVoiceReplyMode(body, cap.reply_mode || 'off', { skipVisibilityToggle: true });
+        // Voice-timbre picker depends on provider+model; rebuilt by callbacks.
+        const modelWrap = body.querySelector(`#cap-${def.id}-model-wrap`);
+        if (modelWrap) {
+            const voiceWrap = document.createElement('div');
+            voiceWrap.id = `cap-${def.id}-voice-wrap`;
+            voiceWrap.innerHTML = `
+                <label class="block text-sm font-medium text-slate-600 dark:text-slate-400 mb-1.5">${t('models_voice')}</label>
+                <div id="cap-${def.id}-voice" class="cfg-dropdown" tabindex="0">
+                    <div class="cfg-dropdown-selected">
+                        <span class="cfg-dropdown-text">--</span>
+                        <i class="fas fa-chevron-down cfg-dropdown-arrow"></i>
+                    </div>
+                    <div class="cfg-dropdown-menu"></div>
+                </div>
+                <div id="cap-${def.id}-voice-custom-wrap" class="hidden mt-2">
+                    <input id="cap-${def.id}-voice-custom" type="text"
+                           class="w-full px-3 py-2 text-sm rounded-md border border-slate-200 dark:border-slate-700
+                                  bg-white dark:bg-slate-800 text-slate-700 dark:text-slate-200
+                                  placeholder:text-slate-400 dark:placeholder:text-slate-500
+                                  focus:outline-none focus:ring-2 focus:ring-primary-500"
+                           placeholder="voice id" />
+                </div>
+            `;
+            modelWrap.parentNode.insertBefore(voiceWrap, modelWrap.nextSibling);
+        }
+    }
+
+    // `body` is still detached from `document`; scope lookups locally.
     const provDd = body.querySelector(`#cap-${def.id}-provider`);
-    // initDropdown's option shape is {value, label}; we strip our private
-    // _configured/_tracked fields before handing it over so the helper stays
-    // generic, then re-attach status decorations afterwards.
+    // Strip private fields before handing to the generic initDropdown helper.
     const ddOpts = providerOpts.map(o => ({ value: o.value, label: o.label }));
 
     let pendingProvider = null;
@@ -3860,15 +4069,9 @@ function renderCapabilityBody(def, cap, body) {
         pendingCapabilitySelection = null;
     }
 
-    // For auto-capable capabilities, an "auto" strategy means the user has
-    // not pinned a vendor; we honor that by selecting the empty-string
-    // sentinel rather than the resolved fallback provider name.
-    // `suggested_provider` is a UI-only preselect (used by embedding & ASR)
-    // when the user has not pinned a vendor yet — purely cosmetic, not
-    // persisted until the user clicks Save.
-    // For "pick or empty" capabilities (no current, no suggestion), we leave
-    // the dropdown unselected and show a muted placeholder so the user is
-    // nudged to pick explicitly.
+    // Auto strategy => leave empty sentinel selected. `suggested_provider`
+    // is a UI-only preselect (not persisted until the user clicks Save).
+    // No current + no suggestion => leave unselected with a placeholder.
     const noSelectionAndNoHint = !cap.current_provider && !cap.suggested_provider;
     const initialProviderValue = pendingProvider
         ? pendingProvider
@@ -3889,20 +4092,82 @@ function renderCapabilityBody(def, cap, body) {
 
     if (def.needsModel) {
         rebuildCapabilityModelDropdown(def, initialProviderValue, cap.current_model || '', body);
-        // Hide the model picker entirely while the capability is in `auto`
-        // mode — there is nothing useful to pin, and the fallback hint
-        // below explains what'll actually run.
+        // Hide model picker in auto mode — fallback hint below covers it.
         setCapabilityModelPickerVisible(def, initialProviderValue !== '' || !capabilitySupportsAuto(def.id), body);
     }
 
+    if (def.id === 'tts') {
+        rebuildCapabilityVoiceDropdown(
+            initialProviderValue,
+            cap.current_voice || '',
+            body,
+            cap.current_model || ''
+        );
+    }
+
     // Inject auto/router-pending hint banners before the action footer.
     renderCapabilityHints(def, cap, body, initialProviderValue);
+
+    if (def.id === 'tts') {
+        _setTtsConfigVisible(body, (cap.reply_mode || 'off') !== 'off');
+    }
 }
 
-// Toggle visibility of the model picker. Used both at first render and
-// whenever the provider dropdown swings between an explicit vendor and the
-// "auto" sentinel. We toggle the wrapper rather than re-rendering so the
-// existing dropdown state survives a round-trip back to a real vendor.
+// TTS reply-policy dropdown (off / voice_if_voice / always). Persists on
+// change. When off, hides the rest of the TTS card.
+function renderVoiceReplyMode(host, currentMode, options) {
+    options = options || {};
+    const opts = [
+        { value: 'off',            label: t('voice_reply_off') },
+        { value: 'voice_if_voice', label: t('voice_reply_if_voice') },
+        { value: 'always',         label: t('voice_reply_always') },
+    ];
+    const wrap = document.createElement('div');
+    wrap.id = 'voice-reply-mode-wrap';
+    wrap.innerHTML = `
+        <label class="block text-sm font-medium text-slate-600 dark:text-slate-400 mb-1.5">${t('voice_reply_mode_label')}</label>
+        <div id="voice-reply-mode-dd" class="cfg-dropdown" tabindex="0">
+            <div class="cfg-dropdown-selected">
+                <span class="cfg-dropdown-text">--</span>
+                <i class="fas fa-chevron-down cfg-dropdown-arrow"></i>
+            </div>
+            <div class="cfg-dropdown-menu"></div>
+        </div>
+    `;
+    host.prepend(wrap);
+
+    const dd = wrap.querySelector('#voice-reply-mode-dd');
+    const valid = ['off', 'voice_if_voice', 'always'];
+    const initial = valid.includes(currentMode) ? currentMode : 'off';
+    if (!options.skipVisibilityToggle) _setTtsConfigVisible(host, initial !== 'off');
+    initDropdown(dd, opts, initial, (mode) => {
+        if (!valid.includes(mode)) return;
+        _setTtsConfigVisible(host, mode !== 'off');
+        fetch('/api/models', {
+            method: 'POST',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({ action: 'set_voice_reply_mode', mode }),
+        })
+            .then(r => r.json())
+            .then(data => {
+                if (data && data.status === 'success') {
+                    _ttsReadyPromise = null;  // force re-probe on next bubble
+                }
+            })
+            .catch(() => {});
+    });
+}
+
+// Show/hide everything in the TTS card below the reply-mode dropdown.
+function _setTtsConfigVisible(host, visible) {
+    if (!host) return;
+    Array.from(host.children).forEach((child) => {
+        if (child.id === 'voice-reply-mode-wrap') return;
+        child.classList.toggle('hidden', !visible);
+    });
+}
+
+// Toggle wrapper visibility instead of re-rendering so dropdown state survives.
 function setCapabilityModelPickerVisible(def, visible, scope) {
     const root = scope || document;
     const wrap = root.querySelector(`#cap-${def.id}-model-wrap`);
@@ -4135,7 +4400,7 @@ function rebuildCapabilityModelDropdown(def, providerId, selectedModel, scope) {
 
     initDropdown(el, opts, initialValue, (value) => {
         const customWrap = document.getElementById(`cap-${def.id}-model-custom-wrap`);
-        if (!customWrap) return;
+        if (customWrap) {
             if (value === '__custom__') {
                 customWrap.classList.remove('hidden');
                 const input = document.getElementById(`cap-${def.id}-model-custom`);
@@ -4143,6 +4408,14 @@ function rebuildCapabilityModelDropdown(def, providerId, selectedModel, scope) {
             } else {
                 customWrap.classList.add('hidden');
             }
+        }
+        // TTS voice catalog may be scoped per engine model (aggregating
+        // gateways). Rebuild the voice picker whenever the model changes.
+        if (def.id === 'tts') {
+            const provDd = document.getElementById('cap-tts-provider');
+            const provId = provDd ? getDropdownValue(provDd) : '';
+            rebuildCapabilityVoiceDropdown(provId, '', null, value);
+        }
     });
 
     const customWrap = root.querySelector(`#cap-${def.id}-model-custom-wrap`);
@@ -4157,22 +4430,93 @@ function rebuildCapabilityModelDropdown(def, providerId, selectedModel, scope) {
     }
 }
 
+// TTS-only: rebuild the voice timbre picker against the provider's
+// curated voice list. Hidden when no provider is picked.
+//
+// Each voice entry may be:
+//   - a bare string  (code = label)
+//   - {value, label, hint?}   so we can show a friendly Chinese name
+//     while persisting the raw API code that the runtime sends.
+function rebuildCapabilityVoiceDropdown(providerId, selectedVoice, scope, modelId) {
+    const root = scope || document;
+    const wrap = root.querySelector(`#cap-tts-voice-wrap`);
+    const el = root.querySelector(`#cap-tts-voice`);
+    if (!wrap || !el) return;
+    const cap = modelsState.capabilities.tts || {};
+    const voicesByProvider = cap.provider_voices || {};
+    let raw = (providerId && voicesByProvider[providerId]) || [];
+    // Some providers (gateways) scope voices by engine model id.
+    if (raw && !Array.isArray(raw) && typeof raw === 'object') {
+        const activeModel = modelId
+            || (root.querySelector(`#cap-tts-model`) ? getDropdownValue(root.querySelector(`#cap-tts-model`)) : '');
+        raw = (activeModel && raw[activeModel]) || [];
+    }
+    if (!raw || raw.length === 0) {
+        wrap.classList.add('hidden');
+        return;
+    }
+    wrap.classList.remove('hidden');
+    // Voice picker: friendly name on the left, raw API code as right-hand
+    // hint. Persisted/sent value is always the raw code.
+    const codes = [];
+    const opts = raw.map(entry => {
+        if (typeof entry === 'string') {
+            codes.push(entry);
+            return { value: entry, label: entry };
+        }
+        codes.push(entry.value);
+        const code = entry.value;
+        const desc = entry.hint || entry.label || code;
+        return {
+            value: code,
+            label: desc,
+            hint: desc === code ? '' : code,
+        };
+    });
+    opts.push({ value: '__custom__', label: currentLang === 'zh' ? '自定义...' : 'Custom...' });
+
+    // Off-catalog values route through the custom branch.
+    let initial = selectedVoice || '';
+    const isCustom = initial && !codes.includes(initial);
+    if (isCustom) initial = '__custom__';
+    if (!initial) initial = codes[0];
+
+    initDropdown(el, opts, initial, (value) => {
+        const customWrap = root.querySelector(`#cap-tts-voice-custom-wrap`);
+        if (!customWrap) return;
+        if (value === '__custom__') {
+            customWrap.classList.remove('hidden');
+            const input = root.querySelector(`#cap-tts-voice-custom`);
+            if (input && !input.value) input.value = isCustom ? selectedVoice : '';
+        } else {
+            customWrap.classList.add('hidden');
+        }
+    });
+
+    const customWrap = root.querySelector(`#cap-tts-voice-custom-wrap`);
+    if (customWrap) {
+        if (initial === '__custom__') {
+            customWrap.classList.remove('hidden');
+            const input = root.querySelector(`#cap-tts-voice-custom`);
+            if (input) input.value = isCustom ? selectedVoice : '';
+        } else {
+            customWrap.classList.add('hidden');
+        }
+    }
+}
+
 function onCapabilityProviderChange(def, providerId, scope) {
     if (def.needsModel) {
-        // For capabilities that support `auto`, switching to the empty
-        // sentinel hides the model picker entirely so the card reads as
-        // "we'll figure it out"; switching back to a real vendor re-runs
-        // the rebuild against the capability-scoped model list.
+        // Empty sentinel hides the model picker (capability is in auto mode).
         const isAuto = providerId === '' && capabilitySupportsAuto(def.id);
         if (!isAuto) {
             rebuildCapabilityModelDropdown(def, providerId, '', scope);
         }
         setCapabilityModelPickerVisible(def, !isAuto, scope);
     }
-    // Refresh the auto-hint so it disappears once the user pins a vendor
-    // and reappears when they swing back to "auto". renderCapabilityHints
-    // now writes directly into the footer's hint slot, so we just call it
-    // again — no need to clean up stale DOM nodes.
+    if (def.id === 'tts') {
+        rebuildCapabilityVoiceDropdown(providerId, '', scope);
+    }
     const body = scope || document.querySelector(`[data-cap-body="${def.id}"]`);
     if (body) {
         const cap = modelsState.capabilities[def.id] || {};
@@ -4202,6 +4546,16 @@ function saveCapability(capId) {
     // the backend treats this as "fall back to the runtime chain".
     const isAuto = provider === '' && capabilitySupportsAuto(capId);
     const model = isAuto ? '' : getCapabilityModelValue(def);
+    // TTS carries an extra voice timbre (supports free-text custom ids).
+    let voice = '';
+    if (capId === 'tts' && !isAuto) {
+        const voiceDd = document.getElementById(`cap-${capId}-voice`);
+        voice = voiceDd ? getDropdownValue(voiceDd) : '';
+        if (voice === '__custom__') {
+            const input = document.getElementById(`cap-${capId}-voice-custom`);
+            voice = input ? input.value.trim() : '';
+        }
+    }
 
     // Embedding changes invalidate any pre-existing vector index because
     // dimensions / vendor differ. Gate the save behind a confirm, and on
@@ -4243,19 +4597,19 @@ function saveCapability(capId) {
             return;
         }
     }
-    _persistCapability(capId, provider, model);
+    _persistCapability(capId, provider, model, undefined, { voice });
 }
 
-function _persistCapability(capId, provider, model, onAfterSuccess) {
+function _persistCapability(capId, provider, model, onAfterSuccess, extras) {
+    const payload = { action: 'set_capability', capability: capId, provider_id: provider, model: model };
+    if (extras && extras.voice !== undefined) payload.voice = extras.voice;
     fetch('/api/models', {
         method: 'POST',
         headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ action: 'set_capability', capability: capId, provider_id: provider, model: model }),
+        body: JSON.stringify(payload),
     }).then(r => r.json()).then(data => {
         if (data.status === 'success') {
-            // Show "Saved" first, then refresh — loadModelsView would
-            // otherwise rebuild the card and wipe the status span before
-            // the user can register the confirmation.
+            // Flash "Saved" before reload so the status survives the rebuild.
             showStatus(`cap-${capId}-status`, 'models_save_success', false);
             setTimeout(() => {
                 loadModelsView({ preserveScroll: true });

channel/web/web_channel.py

@@ -6,6 +6,7 @@ import logging
 import mimetypes
 import os
 import random
+import shutil
 import threading
 import time
 import uuid
@@ -295,6 +296,12 @@ class WebChannel(ChatChannel):
                     "timestamp": time.time()
                 })
                 logger.debug(f"SSE done sent for request {request_id}")
+                # Auto-trigger TTS once the bot finishes its text reply. The
+                # synthesis runs in the background so the chat stream is never
+                # blocked; the resulting audio URL is pushed via a follow-up
+                # `voice_attach` SSE event and persisted to messages.extras.
+                if reply.type == ReplyType.TEXT and content.strip():
+                    self._maybe_dispatch_auto_tts(request_id, session_id, content, context)
                 return
 
             # Fallback: polling mode
@@ -461,16 +468,133 @@ class WebChannel(ChatChannel):
 
         return on_event
 
+    # ------------------------------------------------------------------
+    # TTS auto-dispatch
+    # ------------------------------------------------------------------
+    @staticmethod
+    def _resolve_voice_reply_mode() -> str:
+        """
+        Decide the TTS auto-reply policy.
+
+        Source of truth is the cross-channel pair
+        (`always_reply_voice`, `voice_reply_voice`) which chat_channel
+        also consults. The web UI presents these as a single three-state
+        picker (off / voice_if_voice / always) via a lossless mapping.
+        """
+        if conf().get("always_reply_voice", False):
+            return "always"
+        if conf().get("voice_reply_voice", False):
+            return "voice_if_voice"
+        return "off"
+
+    # Mirror of ModelsHandler._TTS_PROVIDERS. zhipu is intentionally omitted
+    # from the UI (GLM-TTS prelude beep); pinning it in config.json still works.
+    _TTS_PROVIDERS_SUGGEST_ORDER = ["openai", "minimax", "dashscope", "linkai"]
+
+    @classmethod
+    def _tts_provider_ready(cls) -> bool:
+        """True if user picked a provider OR any suggested vendor has an API key."""
+        if (conf().get("text_to_voice") or "").strip():
+            return True
+        for pid in cls._TTS_PROVIDERS_SUGGEST_ORDER:
+            meta = ConfigHandler.PROVIDER_MODELS.get(pid) or {}
+            key_field = meta.get("api_key_field")
+            if not key_field:
+                continue
+            val = (conf().get(key_field) or "").strip()
+            if val and val not in ("YOUR API KEY", "YOUR_API_KEY"):
+                return True
+        return False
+
+    def _maybe_dispatch_auto_tts(
+        self,
+        request_id: str,
+        session_id: str,
+        text: str,
+        context: dict,
+    ) -> None:
+        try:
+            mode = self._resolve_voice_reply_mode()
+            if mode == "off":
+                return
+            if mode == "voice_if_voice" and not context.get("is_voice_input"):
+                return
+            if not self._tts_provider_ready():
+                return
+            threading.Thread(
+                target=self._synthesize_tts_async,
+                args=(request_id, session_id, text),
+                daemon=True,
+            ).start()
+        except Exception as e:
+            logger.debug(f"[WebChannel] auto-tts dispatch skipped: {e}")
+
+    def _synthesize_tts_async(
+        self,
+        request_id: str,
+        session_id: str,
+        text: str,
+    ) -> None:
+        try:
+            from bridge.bridge import Bridge
+            reply = Bridge().fetch_text_to_voice(text)
+            if reply is None or reply.type != ReplyType.VOICE or not reply.content:
+                logger.warning(
+                    f"[WebChannel] TTS produced no audio for request {request_id}: "
+                    f"reply={reply}"
+                )
+                return
+            url = self._publish_tts_audio(reply.content)
+            if not url:
+                logger.warning(f"[WebChannel] TTS publish failed for request {request_id}")
+                return
+            payload = {"audio": {"url": url, "kind": "tts"}}
+            try:
+                from agent.memory import get_conversation_store
+                get_conversation_store().attach_extras_to_last_assistant(session_id, payload)
+            except Exception as e:
+                logger.debug(f"[WebChannel] tts persist skipped: {e}")
+            q = self.sse_queues.get(request_id)
+            if q is None:
+                logger.warning(
+                    f"[WebChannel] TTS ready but SSE queue already closed "
+                    f"for request {request_id} (url={url})"
+                )
+                return
+            q.put({
+                "type": "voice_attach",
+                "url": url,
+                "request_id": request_id,
+                "timestamp": time.time(),
+            })
+            logger.info(f"[WebChannel] TTS voice_attach pushed for request {request_id}: {url}")
+        except Exception as e:
+            # TTS failures are intentionally silent (no user-facing error).
+            logger.warning(f"[WebChannel] TTS synthesis failed: {e}")
+
+    @staticmethod
+    def _publish_tts_audio(src_path: str) -> str:
+        """Move a TTS file into uploads/ and return its public URL."""
+        try:
+            if not src_path or not os.path.isfile(src_path):
+                logger.warning(f"[WebChannel] publish_tts_audio missing source: {src_path!r}")
+                return ""
+            ext = os.path.splitext(src_path)[1].lower() or ".mp3"
+            upload_dir = _get_upload_dir()
+            os.makedirs(upload_dir, exist_ok=True)
+            ts = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
+            dst_name = f"voice_reply_{ts}_{random.randint(0, 9999)}{ext}"
+            dst_path = os.path.join(upload_dir, dst_name)
+            shutil.move(src_path, dst_path)
+            logger.debug(f"[WebChannel] publish_tts_audio moved {src_path} -> {dst_path}")
+            return f"/uploads/{dst_name}"
+        except Exception as e:
+            logger.warning(f"[WebChannel] publish_tts_audio failed: {e}")
+            return ""
+
     @staticmethod
     def _cleanup_stale_voice_recordings(max_age_seconds: int = 3600) -> None:
-        """Delete voice-input audio files older than `max_age_seconds`.
-
-        Called once at startup. Web mic recordings live in the upload
-        directory so the browser can replay them inside the conversation
-        bubble. We don't persist them to history, so once a process
-        restarts they're useless — but they're never auto-cleaned
-        anywhere else, so without this they accumulate over time.
-        """
+        """Drop voice_input_* uploads older than max_age_seconds (run at startup)."""
         try:
             upload_dir = _get_upload_dir()
             if not os.path.isdir(upload_dir):
@@ -619,6 +743,10 @@ class WebChannel(ChatChannel):
             prompt = json_data.get('message', '')
             use_sse = json_data.get('stream', True)
             attachments = json_data.get('attachments', [])
+            # Tag the message as originating from voice input so the post-reply
+            # TTS hook can honour the `voice_if_voice` policy (mirrors the
+            # desire_rtype concept used by other channels).
+            is_voice_input = bool(json_data.get('is_voice', False))
 
             # Append file references to the prompt (same format as QQ channel)
             if attachments:
@@ -669,6 +797,11 @@ class WebChannel(ChatChannel):
             context["session_id"] = session_id
             context["receiver"] = session_id
             context["request_id"] = request_id
+            if is_voice_input:
+                # Web channel runs its own TTS post-pipeline via
+                # _maybe_dispatch_auto_tts; don't set desire_rtype here or
+                # chat_channel would synthesize a duplicate VOICE reply.
+                context["is_voice_input"] = True
 
             if use_sse:
                 context["on_event"] = self._make_sse_callback(request_id)
@@ -696,27 +829,39 @@ class WebChannel(ChatChannel):
         q = self.sse_queues[request_id]
         idle_timeout = 600  # 10 minutes without any real event
         deadline = time.time() + idle_timeout
-        done = False
+        # After the main reply is done we keep the stream open for a short
+        # tail so async post-processing (TTS auto-synthesis) can deliver a
+        # `voice_attach` event before the client disconnects.
+        POST_DONE_TAIL_SECONDS = 60
+        post_done = False
+        post_deadline = 0.0
 
         try:
             while time.time() < deadline:
                 try:
                     item = q.get(timeout=1)
                 except Empty:
+                    if post_done and time.time() >= post_deadline:
+                        break
                     yield b": keepalive\n\n"
                     continue
 
-                # Real event received, reset idle deadline
                 deadline = time.time() + idle_timeout
-
                 payload = json.dumps(item, ensure_ascii=False)
                 yield f"data: {payload}\n\n".encode("utf-8")
 
-                if item.get("type") == "done":
-                    done = True
-                    break
+                itype = item.get("type")
+                if itype == "done":
+                    post_done = True
+                    post_deadline = time.time() + POST_DONE_TAIL_SECONDS
+                elif itype == "voice_attach":
+                    # WSGI buffers the previous chunk until the next yield;
+                    # shrink the tail so the generator wakes up quickly to
+                    # emit a couple of keepalive comments that push the
+                    # voice_attach payload through to the browser.
+                    post_done = True
+                    post_deadline = time.time() + 2  # 2s post-attach tail
         finally:
-            if done:
             self.sse_queues.pop(request_id, None)
 
     def poll_response(self):
@@ -811,6 +956,7 @@ class WebChannel(ChatChannel):
             '/uploads/(.*)', 'UploadsHandler',
             '/api/file', 'FileServeHandler',
             '/api/voice/asr', 'VoiceAsrHandler',
+            '/api/voice/tts', 'VoiceTtsHandler',
             '/poll', 'PollHandler',
             '/stream', 'StreamHandler',
             '/chat', 'ChatHandler',
@@ -936,15 +1082,8 @@ class UploadHandler:
 
 
 class VoiceAsrHandler:
-    """
-    Accept a short audio recording from the web console mic button,
-    save it under uploads/ so the browser can replay it, then run it
-    through the currently configured ASR provider.
-
-    Returns {status, text, audio_url} on success — the frontend renders
-    a voice-message bubble with the playable audio and the transcribed
-    caption.
-    """
+    """Receive a mic recording, persist it under uploads/ and run ASR.
+    Returns {status, text, audio_url} so the UI can render a playback bubble."""
     def POST(self):
         _require_auth()
         web.header('Content-Type', 'application/json; charset=utf-8')
@@ -997,6 +1136,48 @@ class VoiceAsrHandler:
             return json.dumps({"status": "error", "message": str(e)})
 
 
+class VoiceTtsHandler:
+    """On-demand TTS for the in-chat "read aloud" button. Returns the
+    audio URL and (when session_id is given) persists it onto the message."""
+    def POST(self):
+        _require_auth()
+        web.header('Content-Type', 'application/json; charset=utf-8')
+        try:
+            data = json.loads(web.data() or b"{}")
+            text = (data.get("text") or "").strip()
+            session_id = (data.get("session_id") or "").strip()
+            if not text:
+                return json.dumps({"status": "error", "message": "empty text"})
+            # `@singleton` makes WebChannel a factory function — go via instance.
+            channel = WebChannel()
+            if not channel._tts_provider_ready():
+                return json.dumps({"status": "error", "message": "tts not configured"})
+
+            from bridge.bridge import Bridge
+            reply = Bridge().fetch_text_to_voice(text)
+            if reply is None or reply.type != ReplyType.VOICE or not reply.content:
+                msg = getattr(reply, "content", "") or "tts failed"
+                return json.dumps({"status": "error", "message": str(msg)})
+
+            url = channel._publish_tts_audio(reply.content)
+            if not url:
+                return json.dumps({"status": "error", "message": "publish failed"})
+
+            if session_id:
+                try:
+                    from agent.memory import get_conversation_store
+                    get_conversation_store().attach_extras_to_last_assistant(
+                        session_id, {"audio": {"url": url, "kind": "tts"}},
+                    )
+                except Exception as e:
+                    logger.debug(f"[VoiceTtsHandler] persist skipped: {e}")
+
+            return json.dumps({"status": "success", "audio_url": url})
+        except Exception as e:
+            logger.exception(f"[VoiceTtsHandler] failed: {e}")
+            return json.dumps({"status": "error", "message": str(e)})
+
+
 class UploadsHandler:
     def GET(self, file_name):
         _require_auth()
@@ -1357,10 +1538,243 @@ class ModelsHandler:
     POST /api/models/capability -> set provider/model for a capability
     """
 
-    # Capability -> editable flag, current-value resolver, and supported provider
-    # ids drawn from ConfigHandler.PROVIDER_MODELS where applicable.
+    # Capability -> provider ids drawn from ConfigHandler.PROVIDER_MODELS.
     _ASR_PROVIDERS = ["openai", "dashscope", "zhipu", "linkai"]
-    _TTS_PROVIDERS = ["openai", "linkai", "minimax", "baidu", "ali", "xunfei", "azure", "google", "elevenlabs", "edge", "pytts"]
+    # Web-console white-list. Other vendors stay usable via direct config.
+    _TTS_PROVIDERS = ["openai", "minimax", "dashscope", "linkai"]
+
+    # TTS engine catalog (speech models, not voice timbres). Entries are
+    # either a bare code or {value, hint?} when a friendly label helps.
+    _TTS_PROVIDER_MODELS = {
+        "openai":    ["tts-1", "tts-1-hd", "gpt-4o-mini-tts"],
+        "minimax": [
+            {"value": "speech-2.8-hd",    "hint": "情绪渲染融合语气词,自然听感"},
+            {"value": "speech-2.8-turbo", "hint": "极致生成速度,更自然逼真"},
+            {"value": "speech-2.6-hd",    "hint": "超低延时,归一化升级"},
+            {"value": "speech-2.6-turbo", "hint": "更快更便宜,适合语音聊天/数字人"},
+        ],
+        "dashscope": [
+            {"value": "qwen3-tts-flash", "hint": "覆盖普通话、方言与主流外语"},
+        ],
+        # Aggregating gateway: a single endpoint multiplexes several
+        # underlying TTS engines, selected via the `model` field.
+        # Each engine exposes its own voice catalog (see _TTS_PROVIDER_VOICES).
+        "linkai": [
+            {"value": "tts-1",  "hint": "OpenAI · 多语种通用"},
+            {"value": "doubao", "hint": "字节豆包 · 中文音色丰富"},
+            {"value": "baidu",  "hint": "百度 · 中文主播音色"},
+        ],
+    }
+
+    # Per-provider voice timbres. Entries can be a bare code string
+    # (label = code) or {value, hint?} when a friendly secondary label
+    # helps recognition. We keep `value` as the raw API code so power
+    # users can cross-reference config.json.
+    _TTS_PROVIDER_VOICES = {
+        "openai":    [
+            "alloy", "echo", "fable", "onyx", "nova", "shimmer",
+            "ash", "ballad", "coral", "sage", "verse",
+        ],
+        "minimax": [
+            # Mandarin Chinese (full catalog)
+            {"value": "male-qn-qingse",                           "hint": "中文 · 青涩青年（男）"},
+            {"value": "male-qn-jingying",                         "hint": "中文 · 精英青年（男）"},
+            {"value": "male-qn-badao",                            "hint": "中文 · 霸道青年（男）"},
+            {"value": "male-qn-daxuesheng",                       "hint": "中文 · 青年大学生（男）"},
+            {"value": "female-shaonv",                            "hint": "中文 · 少女（女）"},
+            {"value": "female-yujie",                             "hint": "中文 · 御姐（女）"},
+            {"value": "female-chengshu",                          "hint": "中文 · 成熟女性（女）"},
+            {"value": "female-tianmei",                           "hint": "中文 · 甜美女性（女）"},
+            {"value": "male-qn-qingse-jingpin",                   "hint": "中文 · 青涩青年-beta（男）"},
+            {"value": "male-qn-jingying-jingpin",                 "hint": "中文 · 精英青年-beta（男）"},
+            {"value": "male-qn-badao-jingpin",                    "hint": "中文 · 霸道青年-beta（男）"},
+            {"value": "male-qn-daxuesheng-jingpin",               "hint": "中文 · 青年大学生-beta（男）"},
+            {"value": "female-shaonv-jingpin",                    "hint": "中文 · 少女-beta（女）"},
+            {"value": "female-yujie-jingpin",                     "hint": "中文 · 御姐-beta（女）"},
+            {"value": "female-chengshu-jingpin",                  "hint": "中文 · 成熟女性-beta（女）"},
+            {"value": "female-tianmei-jingpin",                   "hint": "中文 · 甜美女性-beta（女）"},
+            {"value": "clever_boy",                               "hint": "中文 · 聪明男童"},
+            {"value": "cute_boy",                                 "hint": "中文 · 可爱男童"},
+            {"value": "lovely_girl",                              "hint": "中文 · 萌萌女童"},
+            {"value": "cartoon_pig",                              "hint": "中文 · 卡通猪小琪"},
+            {"value": "bingjiao_didi",                            "hint": "中文 · 病娇弟弟"},
+            {"value": "junlang_nanyou",                           "hint": "中文 · 俊朗男友"},
+            {"value": "chunzhen_xuedi",                           "hint": "中文 · 纯真学弟"},
+            {"value": "lengdan_xiongzhang",                       "hint": "中文 · 冷淡学长"},
+            {"value": "badao_shaoye",                             "hint": "中文 · 霸道少爷"},
+            {"value": "tianxin_xiaoling",                         "hint": "中文 · 甜心小玲"},
+            {"value": "qiaopi_mengmei",                           "hint": "中文 · 俏皮萌妹"},
+            {"value": "wumei_yujie",                              "hint": "中文 · 妩媚御姐"},
+            {"value": "diadia_xuemei",                            "hint": "中文 · 嗲嗲学妹"},
+            {"value": "danya_xuejie",                             "hint": "中文 · 淡雅学姐"},
+            {"value": "Chinese (Mandarin)_Reliable_Executive",    "hint": "中文 · 沉稳高管"},
+            {"value": "Chinese (Mandarin)_News_Anchor",           "hint": "中文 · 新闻女声"},
+            {"value": "Chinese (Mandarin)_Mature_Woman",          "hint": "中文 · 傲娇御姐"},
+            {"value": "Chinese (Mandarin)_Unrestrained_Young_Man","hint": "中文 · 不羁青年"},
+            {"value": "Arrogant_Miss",                            "hint": "中文 · 嚣张小姐"},
+            {"value": "Robot_Armor",                              "hint": "中文 · 机械战甲"},
+            {"value": "Chinese (Mandarin)_Kind-hearted_Antie",    "hint": "中文 · 热心大婶"},
+            {"value": "Chinese (Mandarin)_HK_Flight_Attendant",   "hint": "中文 · 港普空姐"},
+            {"value": "Chinese (Mandarin)_Humorous_Elder",        "hint": "中文 · 搞笑大爷"},
+            {"value": "Chinese (Mandarin)_Gentleman",             "hint": "中文 · 温润男声"},
+            {"value": "Chinese (Mandarin)_Warm_Bestie",           "hint": "中文 · 温暖闺蜜"},
+            {"value": "Chinese (Mandarin)_Male_Announcer",        "hint": "中文 · 播报男声"},
+            {"value": "Chinese (Mandarin)_Sweet_Lady",            "hint": "中文 · 甜美女声"},
+            {"value": "Chinese (Mandarin)_Southern_Young_Man",    "hint": "中文 · 南方小哥"},
+            {"value": "Chinese (Mandarin)_Wise_Women",            "hint": "中文 · 阅历姐姐"},
+            {"value": "Chinese (Mandarin)_Gentle_Youth",          "hint": "中文 · 温润青年"},
+            {"value": "Chinese (Mandarin)_Warm_Girl",             "hint": "中文 · 温暖少女"},
+            {"value": "Chinese (Mandarin)_Kind-hearted_Elder",    "hint": "中文 · 花甲奶奶"},
+            {"value": "Chinese (Mandarin)_Cute_Spirit",           "hint": "中文 · 憨憨萌兽"},
+            {"value": "Chinese (Mandarin)_Radio_Host",            "hint": "中文 · 电台男主播"},
+            {"value": "Chinese (Mandarin)_Lyrical_Voice",         "hint": "中文 · 抒情男声"},
+            {"value": "Chinese (Mandarin)_Straightforward_Boy",   "hint": "中文 · 率真弟弟"},
+            {"value": "Chinese (Mandarin)_Sincere_Adult",         "hint": "中文 · 真诚青年"},
+            {"value": "Chinese (Mandarin)_Gentle_Senior",         "hint": "中文 · 温柔学姐"},
+            {"value": "Chinese (Mandarin)_Stubborn_Friend",       "hint": "中文 · 嘴硬竹马"},
+            {"value": "Chinese (Mandarin)_Crisp_Girl",            "hint": "中文 · 清脆少女"},
+            {"value": "Chinese (Mandarin)_Pure-hearted_Boy",      "hint": "中文 · 清澈邻家弟弟"},
+            {"value": "Chinese (Mandarin)_Soft_Girl",             "hint": "中文 · 柔和少女"},
+            # Cantonese (full catalog)
+            {"value": "Cantonese_ProfessionalHost（F)",            "hint": "粤语 · 专业女主持"},
+            {"value": "Cantonese_GentleLady",                     "hint": "粤语 · 温柔女声"},
+            {"value": "Cantonese_ProfessionalHost（M)",            "hint": "粤语 · 专业男主持"},
+            {"value": "Cantonese_PlayfulMan",                     "hint": "粤语 · 活泼男声"},
+            {"value": "Cantonese_CuteGirl",                       "hint": "粤语 · 可爱女孩"},
+            {"value": "Cantonese_KindWoman",                      "hint": "粤语 · 善良女声"},
+            # English (curated: 1F + 1M)
+            {"value": "English_Graceful_Lady",                    "hint": "英文 · Graceful Lady（女）"},
+            {"value": "English_Trustworthy_Man",                  "hint": "英文 · Trustworthy Man（男）"},
+            # Japanese (curated: 1F + 1M)
+            {"value": "Japanese_KindLady",                        "hint": "日文 · Kind Lady（女）"},
+            {"value": "Japanese_LoyalKnight",                     "hint": "日文 · Loyal Knight（男）"},
+            # Korean (curated: 1F + 1M)
+            {"value": "Korean_SweetGirl",                         "hint": "韩文 · Sweet Girl（女）"},
+            {"value": "Korean_CheerfulBoyfriend",                 "hint": "韩文 · Cheerful Boyfriend（男）"},
+        ],
+        "dashscope": [
+            {"value": "Cherry",   "hint": "芊悦 · 阳光女声"},
+            {"value": "Serena",   "hint": "苏瑶 · 温柔女声"},
+            {"value": "Chelsie",  "hint": "千雪 · 二次元少女"},
+            {"value": "Ethan",    "hint": "晨煦 · 阳光男声"},
+            {"value": "Moon",     "hint": "月白 · 率性男声"},
+            {"value": "Kai",      "hint": "凯 · 治愈男声"},
+            {"value": "Nofish",   "hint": "不吃鱼 · 设计师男声"},
+            {"value": "Bella",    "hint": "萌宝 · 小萝莉"},
+            {"value": "Bunny",    "hint": "萌小姬 · 萌系少女"},
+            {"value": "Stella",   "hint": "少女阿月 · 元气少女"},
+            {"value": "Neil",     "hint": "阿闻 · 新闻主播"},
+            {"value": "Seren",    "hint": "小婉 · 助眠女声"},
+            {"value": "Jada",     "hint": "上海话 · 阿珍"},
+            {"value": "Dylan",    "hint": "北京话 · 晓东"},
+            {"value": "Sunny",    "hint": "四川话 · 晴儿"},
+            {"value": "Eric",     "hint": "四川话 · 程川"},
+            {"value": "Rocky",    "hint": "粤语 · 阿强"},
+            {"value": "Kiki",     "hint": "粤语 · 阿清"},
+            {"value": "Peter",    "hint": "天津话 · 李彼得"},
+            {"value": "Marcus",   "hint": "陕西话 · 秦川"},
+            {"value": "Roy",      "hint": "闽南语 · 阿杰"},
+        ],
+        # Aggregating gateway: voices are scoped per engine model. The
+        # frontend picks the correct list based on the selected model so
+        # users don't see incompatible timbres for the active engine.
+        "linkai": {
+            "tts-1": [
+                "alloy", "echo", "fable", "onyx", "nova", "shimmer",
+            ],
+            "doubao": [
+                {"value": "zh_female_wanwanxiaohe_moon_bigtts",       "hint": "湾湾小何"},
+                {"value": "BV007_streaming",                          "hint": "亲切女声"},
+                {"value": "BV001_streaming",                          "hint": "通用女声"},
+                {"value": "BV002_streaming",                          "hint": "通用男声"},
+                {"value": "BV051_streaming",                          "hint": "奶气萌娃"},
+                {"value": "zh_female_linjianvhai_moon_bigtts",        "hint": "邻家女孩"},
+                {"value": "BV700_streaming",                          "hint": "灿灿"},
+                {"value": "BV019_streaming",                          "hint": "重庆小伙"},
+                {"value": "BV524_streaming",                          "hint": "日语男声"},
+                {"value": "BV021_streaming",                          "hint": "东北老铁"},
+                {"value": "BV701_streaming",                          "hint": "擎苍"},
+                {"value": "BV113_streaming",                          "hint": "甜宠少御"},
+                {"value": "BV056_streaming",                          "hint": "阳光男声"},
+                {"value": "BV213_streaming",                          "hint": "广西表哥"},
+                {"value": "BV119_streaming",                          "hint": "通用赘婿"},
+                {"value": "BV705_streaming",                          "hint": "炀炀"},
+                {"value": "BV033_streaming",                          "hint": "温柔小哥"},
+                {"value": "BV102_streaming",                          "hint": "儒雅青年"},
+                {"value": "BV522_streaming",                          "hint": "气质女生"},
+                {"value": "BV034_streaming",                          "hint": "知性姐姐 · 双语"},
+                {"value": "BV005_streaming",                          "hint": "活泼女声"},
+                {"value": "zh_female_wanqudashu_moon_bigtts",         "hint": "湾区大叔"},
+                {"value": "zh_female_daimengchuanmei_moon_bigtts",    "hint": "呆萌川妹"},
+                {"value": "zh_male_guozhoudege_moon_bigtts",          "hint": "广州德哥"},
+                {"value": "zh_male_beijingxiaoye_moon_bigtts",        "hint": "北京小爷"},
+                {"value": "zh_male_shaonianzixin_moon_bigtts",        "hint": "少年梓辛 / Brayan"},
+                {"value": "zh_female_meilinvyou_moon_bigtts",         "hint": "魅力女友"},
+                {"value": "zh_male_shenyeboke_moon_bigtts",           "hint": "深夜播客"},
+                {"value": "zh_female_sajiaonvyou_moon_bigtts",        "hint": "柔美女友"},
+                {"value": "zh_female_yuanqinvyou_moon_bigtts",        "hint": "撒娇学妹"},
+                {"value": "zh_male_haoyuxiaoge_moon_bigtts",          "hint": "浩宇小哥"},
+                {"value": "zh_male_guangxiyuanzhou_moon_bigtts",      "hint": "广西远舟"},
+                {"value": "zh_female_meituojieer_moon_bigtts",        "hint": "妹坨洁儿"},
+                {"value": "zh_male_yuzhouzixuan_moon_bigtts",         "hint": "豫州子轩"},
+                {"value": "BV115_streaming",                          "hint": "古风少御"},
+                {"value": "zh_female_gaolengyujie_moon_bigtts",       "hint": "高冷御姐"},
+                {"value": "zh_male_yuanboxiaoshu_moon_bigtts",        "hint": "渊博小叔"},
+                {"value": "zh_male_yangguangqingnian_moon_bigtts",    "hint": "阳光青年"},
+                {"value": "zh_male_aojiaobazong_moon_bigtts",         "hint": "傲娇霸总"},
+                {"value": "zh_male_jingqiangkanye_moon_bigtts",       "hint": "京腔侃爷 / Harmony"},
+                {"value": "zh_female_shuangkuaisisi_moon_bigtts",     "hint": "爽快思思 / Skye"},
+                {"value": "zh_male_wennuanahu_moon_bigtts",           "hint": "温暖阿虎 / Alvin"},
+                {"value": "multi_female_shuangkuaisisi_moon_bigtts",  "hint": "はるこ / Esmeralda"},
+                {"value": "multi_male_jingqiangkanye_moon_bigtts",    "hint": "かずね / Javier or Álvaro"},
+                {"value": "multi_female_gaolengyujie_moon_bigtts",    "hint": "あけみ"},
+                {"value": "multi_male_wanqudashu_moon_bigtts",        "hint": "ひろし / Roberto"},
+                {"value": "ICL_zh_female_bingruoshaonv_tob",          "hint": "病弱少女"},
+                {"value": "ICL_zh_female_huoponvhai_tob",             "hint": "活泼女孩"},
+                {"value": "ICL_zh_female_heainainai_tob",             "hint": "和蔼奶奶"},
+                {"value": "ICL_zh_female_linjuayi_tob",               "hint": "邻居阿姨"},
+                {"value": "zh_female_wenrouxiaoya_moon_bigtts",       "hint": "温柔小雅"},
+                {"value": "zh_female_tianmeixiaoyuan_moon_bigtts",    "hint": "甜美小源"},
+                {"value": "zh_female_qingchezizi_moon_bigtts",        "hint": "清澈梓梓"},
+                {"value": "zh_male_dongfanghaoran_moon_bigtts",       "hint": "东方浩然"},
+                {"value": "zh_male_jieshuoxiaoming_moon_bigtts",      "hint": "解说小明"},
+                {"value": "zh_female_kailangjiejie_moon_bigtts",      "hint": "开朗姐姐"},
+                {"value": "zh_male_linjiananhai_moon_bigtts",         "hint": "邻家男孩"},
+                {"value": "zh_female_tianmeiyueyue_moon_bigtts",      "hint": "甜美悦悦"},
+                {"value": "zh_female_xinlingjitang_moon_bigtts",      "hint": "心灵鸡汤"},
+            ],
+            "baidu": [
+                {"value": "baidu_0",    "hint": "度小美 · 标准女主播"},
+                {"value": "baidu_1",    "hint": "度小宇 · 亲切男声"},
+                {"value": "baidu_3",    "hint": "度逍遥 · 情感男声"},
+                {"value": "baidu_4",    "hint": "度丫丫 · 童声"},
+                {"value": "baidu_5",    "hint": "度小娇 · 成熟女主播"},
+                {"value": "baidu_5003", "hint": "度逍遥 · 情感男声"},
+                {"value": "baidu_5118", "hint": "度小鹿 · 甜美女声"},
+                {"value": "baidu_103",  "hint": "度米朵 · 可爱童声"},
+                {"value": "baidu_106",  "hint": "度博文 · 专业男主播"},
+                {"value": "baidu_110",  "hint": "度小童 · 童声主播"},
+                {"value": "baidu_111",  "hint": "度小萌 · 软萌妹子"},
+                {"value": "baidu_4003", "hint": "度逍遥 · 情感男声"},
+                {"value": "baidu_4100", "hint": "度小雯 · 活力女主播"},
+                {"value": "baidu_4103", "hint": "度米朵 · 可爱女声"},
+                {"value": "baidu_4105", "hint": "度灵儿 · 清澈女声"},
+                {"value": "baidu_4106", "hint": "度博文 · 专业男主播"},
+                {"value": "baidu_4115", "hint": "度小贤 · 电台男主播"},
+                {"value": "baidu_4117", "hint": "度小乔 · 活泼女声"},
+                {"value": "baidu_4119", "hint": "度小鹿 · 甜美女声"},
+                {"value": "baidu_4129", "hint": "度小彦 · 知识男主播"},
+                {"value": "baidu_4140", "hint": "度小新 · 专业女主播"},
+                {"value": "baidu_4143", "hint": "度清风 · 配音男声"},
+                {"value": "baidu_4144", "hint": "度姗姗 · 娱乐女声"},
+                {"value": "baidu_4149", "hint": "度星河 · 广告男声"},
+                {"value": "baidu_4206", "hint": "度博文 · 综艺男声"},
+                {"value": "baidu_4226", "hint": "南方 · 电台女主播"},
+                {"value": "baidu_4254", "hint": "度小清 · 广告女声"},
+                {"value": "baidu_4278", "hint": "度小贝 · 知识女主播"},
+            ],
+        },
+    }
     _EMBEDDING_PROVIDERS = ["openai", "dashscope", "doubao", "zhipu", "linkai"]
 
     # Capability-scoped model catalogs. The chat dropdown can reuse the
@@ -1525,7 +1939,7 @@ class ModelsHandler:
     @classmethod
     def _predict_vision_auto(cls, local_config: dict) -> dict:
         """Predict which provider vision.py will actually dispatch to when
-        no tool.vision.model is set. Mirrors the fallback order in
+        no tools.vision.model is set. Mirrors the fallback order in
         agent/tools/vision/vision.py::_resolve_providers so the UI hint
         matches reality."""
         chat = cls._chat_capability(local_config)
@@ -1590,12 +2004,12 @@ class ModelsHandler:
 
     @classmethod
     def _vision_capability(cls, local_config: dict) -> dict:
-        """Vision model. tool.vision.model is the explicit override; otherwise
+        """Vision model. tools.vision.model is the explicit override; otherwise
         the runtime fallback chain in agent/tools/vision/vision.py decides."""
-        tool_conf = local_config.get("tool") or {}
-        if not isinstance(tool_conf, dict):
-            tool_conf = {}
-        vision_conf = tool_conf.get("vision") or {}
+        tools_conf = local_config.get("tools") or local_config.get("tool") or {}
+        if not isinstance(tools_conf, dict):
+            tools_conf = {}
+        vision_conf = tools_conf.get("vision") or {}
         if not isinstance(vision_conf, dict):
             vision_conf = {}
         user_specified = (vision_conf.get("model") or "").strip()
@@ -1652,14 +2066,38 @@ class ModelsHandler:
 
     @classmethod
     def _tts_capability(cls, local_config: dict) -> dict:
-        provider_id = (local_config.get("text_to_voice") or "openai").strip().lower()
+        explicit = (local_config.get("text_to_voice") or "").strip().lower()
+        # Providers outside the white-list don't drive the picker, but their
+        # underlying runtime config is preserved so bridge still routes them.
+        ui_provider = explicit if explicit in cls._TTS_PROVIDERS else ""
+        suggested = ""
+        if not ui_provider:
+            for pid in cls._TTS_PROVIDERS:
+                meta = ConfigHandler.PROVIDER_MODELS.get(pid) or {}
+                key_field = meta.get("api_key_field")
+                if key_field and cls._is_real_key(local_config.get(key_field, "")):
+                    suggested = pid
+                    break
         return {
             "editable": True,
-            "current_provider": provider_id,
-            "current_model": local_config.get("text_to_voice_model", "") or "",
+            "current_provider": ui_provider,
+            "suggested_provider": suggested,
+            "current_model": (local_config.get("text_to_voice_model") or "") if ui_provider else "",
+            "current_voice": (local_config.get("tts_voice_id") or "") if ui_provider else "",
             "providers": cls._TTS_PROVIDERS,
+            "provider_models": cls._TTS_PROVIDER_MODELS,
+            "provider_voices": cls._TTS_PROVIDER_VOICES,
+            "reply_mode": cls._tts_reply_mode(local_config),
         }
 
+    @staticmethod
+    def _tts_reply_mode(local_config: dict) -> str:
+        if local_config.get("always_reply_voice", False):
+            return "always"
+        if local_config.get("voice_reply_voice", False):
+            return "voice_if_voice"
+        return "off"
+
     @classmethod
     def _embedding_capability(cls, local_config: dict) -> dict:
         # Embedding is "pick or empty" — runtime's legacy openai/linkai
@@ -1728,17 +2166,20 @@ class ModelsHandler:
 
     @classmethod
     def _image_capability(cls, local_config: dict) -> dict:
-        """Image generation. Source of truth: config["skill"]["image-generation"]["model"]
+        """Image generation. Source of truth: config["skills"]["image-generation"]["model"]
         (mirrors the per-skill config schema documented in skills/image-generation).
         The runtime resolver in skills/image-generation/scripts/generate.py
         reads this via the SKILL_IMAGE_GENERATION_MODEL env var that the
         agent_initializer syncs at startup; provider is inferred from the
         model name prefix, mirroring vision.py's design.
+
+        ``skill`` (singular) is still tolerated as a legacy fallback —
+        config.load_config() folds it into ``skills`` at startup.
         """
-        skill_node = local_config.get("skill") or {}
-        if not isinstance(skill_node, dict):
-            skill_node = {}
-        img_node = skill_node.get("image-generation") or {}
+        skills_node = local_config.get("skills") or local_config.get("skill") or {}
+        if not isinstance(skills_node, dict):
+            skills_node = {}
+        img_node = skills_node.get("image-generation") or {}
         if not isinstance(img_node, dict):
             img_node = {}
         explicit_model = (img_node.get("model") or "").strip()
@@ -1832,6 +2273,8 @@ class ModelsHandler:
                 return self._handle_delete_provider(data)
             if action == "set_capability":
                 return self._handle_set_capability(data)
+            if action == "set_voice_reply_mode":
+                return self._handle_set_voice_reply_mode(data)
             return json.dumps({"status": "error", "message": f"unknown action: {action!r}"})
         except Exception as e:
             logger.error(f"[ModelsHandler] POST failed: {e}")
@@ -1918,7 +2361,7 @@ class ModelsHandler:
         if capability == "asr":
             return self._set_simple("voice_to_text", provider_id)
         if capability == "tts":
-            return self._set_tts(provider_id, model)
+            return self._set_tts(provider_id, model, (data.get("voice") or "").strip())
         if capability == "embedding":
             return self._set_embedding(provider_id, model)
         if capability == "image":
@@ -1926,35 +2369,20 @@ class ModelsHandler:
         return json.dumps({"status": "error", "message": f"capability not editable: {capability}"})
 
     def _set_image(self, provider_id: str, model: str) -> str:
-        # Source of truth: config["skill"]["image-generation"]["model"].
-        # provider_id is informational only (used by the UI to highlight a
-        # vendor card); the runtime resolver infers the provider from the
-        # model name prefix at request time, mirroring vision.py's design.
-        # An empty model means "switch back to auto / let the script pick".
+        # Source of truth: skills.image-generation.model. provider_id is
+        # informational only; the resolver picks the vendor by model prefix.
         local_config = conf()
         file_cfg = self._read_file_config()
 
-        def _ensure_skill_node(cfg: dict) -> dict:
-            skill_node = cfg.get("skill") or {}
-            if not isinstance(skill_node, dict):
-                skill_node = {}
-            img_node = skill_node.get("image-generation") or {}
-            if not isinstance(img_node, dict):
-                img_node = {}
-            skill_node["image-generation"] = img_node
-            cfg["skill"] = skill_node
-            return img_node
-
-        _ensure_skill_node(local_config)["model"] = model or ""
-        _ensure_skill_node(file_cfg)["model"] = model or ""
+        self._set_nested_namespace_value(local_config, "skills", "image-generation", "model", model or "")
+        self._set_nested_namespace_value(file_cfg, "skills", "image-generation", "model", model or "")
+        self._drop_legacy_namespace(local_config, "skill", "skills", child="image-generation")
+        self._drop_legacy_namespace(file_cfg, "skill", "skills", child="image-generation")
 
         self._write_file_config(file_cfg)
 
-        # The skill subprocess (skills/image-generation/scripts/generate.py)
-        # reads SKILL_IMAGE_GENERATION_MODEL from its environment, which is
-        # only synced from config["skill"] at startup. Update os.environ live
-        # so changes take effect on the next call without a restart. An empty
-        # model means "clear the override" → drop the env var entirely.
+        # The skill subprocess reads SKILL_IMAGE_GENERATION_MODEL from env at
+        # startup; mirror the change so live edits apply without restart.
         env_key = "SKILL_IMAGE_GENERATION_MODEL"
         if model:
             os.environ[env_key] = model
@@ -1992,8 +2420,6 @@ class ModelsHandler:
             applied["model"] = model
 
         if not applied:
-            # No-op save (nothing to write). Return success so the UI can
-            # confirm the click without showing a misleading error.
             return json.dumps({"status": "success", "applied": {}, "noop": True})
 
         self._write_file_config(file_cfg)
@@ -2002,34 +2428,66 @@ class ModelsHandler:
         return json.dumps({"status": "success", "applied": applied})
 
     def _set_vision(self, provider_id: str, model: str) -> str:
-        # Vision uses tool.vision.model (nested). provider_id is informational
-        # only; the runtime resolver auto-routes by model name prefix.
+        # Source of truth: tools.vision.model. provider_id is informational
+        # only; the resolver picks the vendor by model prefix.
         local_config = conf()
         file_cfg = self._read_file_config()
-        tool_node = file_cfg.get("tool") or {}
-        if not isinstance(tool_node, dict):
-            tool_node = {}
-        vision_node = tool_node.get("vision") or {}
-        if not isinstance(vision_node, dict):
-            vision_node = {}
-        vision_node["model"] = model
-        tool_node["vision"] = vision_node
-        file_cfg["tool"] = tool_node
-        # Mirror into in-memory config so the live agent sees the change.
-        runtime_tool = local_config.get("tool") or {}
-        if not isinstance(runtime_tool, dict):
-            runtime_tool = {}
-        runtime_vision = runtime_tool.get("vision") or {}
-        if not isinstance(runtime_vision, dict):
-            runtime_vision = {}
-        runtime_vision["model"] = model
-        runtime_tool["vision"] = runtime_vision
-        local_config["tool"] = runtime_tool
+        self._set_nested_namespace_value(file_cfg, "tools", "vision", "model", model)
+        self._set_nested_namespace_value(local_config, "tools", "vision", "model", model)
+        self._drop_legacy_namespace(file_cfg, "tool", "tools", child="vision")
+        self._drop_legacy_namespace(local_config, "tool", "tools", child="vision")
 
         self._write_file_config(file_cfg)
         logger.info(f"[ModelsHandler] vision model set: {model!r}")
         return json.dumps({"status": "success", "model": model})
 
+    @staticmethod
+    def _set_nested_namespace_value(cfg, top: str, name: str, key: str, value):
+        """Set ``cfg[top][name][key] = value``, creating missing dicts."""
+        bucket = cfg.get(top)
+        if not isinstance(bucket, dict):
+            bucket = {}
+        node = bucket.get(name)
+        if not isinstance(node, dict):
+            node = {}
+        node[key] = value
+        bucket[name] = node
+        cfg[top] = bucket
+
+    @staticmethod
+    def _drop_legacy_namespace(cfg, legacy: str, canonical: str, child: str) -> None:
+        """Strip the deprecated singular key so config.json stays single-source."""
+        legacy_section = cfg.get(legacy)
+        if not isinstance(legacy_section, dict):
+            return
+        legacy_section.pop(child, None)
+        if legacy_section:
+            cfg[legacy] = legacy_section
+        else:
+            cfg.pop(legacy, None)
+
+    def _handle_set_voice_reply_mode(self, data: dict) -> str:
+        # UI picker (off / voice_if_voice / always) maps to the legacy
+        # always_reply_voice + voice_reply_voice pair that chat_channel.py
+        # reads, so all channels (web/feishu/wecom/...) share the routing.
+        mode = (data.get("mode") or "").strip().lower()
+        if mode not in ("off", "voice_if_voice", "always"):
+            return json.dumps({"status": "error", "message": f"invalid mode: {mode!r}"})
+        always = (mode == "always")
+        if_voice = (mode == "voice_if_voice")
+        local_config = conf()
+        file_cfg = self._read_file_config()
+        local_config["always_reply_voice"] = always
+        local_config["voice_reply_voice"] = if_voice
+        file_cfg["always_reply_voice"] = always
+        file_cfg["voice_reply_voice"] = if_voice
+        self._write_file_config(file_cfg)
+        logger.info(
+            f"[ModelsHandler] voice reply mode set: {mode!r} "
+            f"(always_reply_voice={always}, voice_reply_voice={if_voice})"
+        )
+        return json.dumps({"status": "success", "mode": mode})
+
     def _set_simple(self, key: str, value: str) -> str:
         local_config = conf()
         file_cfg = self._read_file_config()
@@ -2037,25 +2495,30 @@ class ModelsHandler:
         file_cfg[key] = value
         self._write_file_config(file_cfg)
         logger.info(f"[ModelsHandler] {key} set: {value!r}")
-        # Bridge caches voice_to_text routing + bot instance; refresh it
-        # so the change takes effect on the next voice request.
+        # Hot-swap the cached voice bot so the change takes effect immediately.
         if key in ("voice_to_text", "text_to_voice"):
             self._refresh_voice_routing()
         return json.dumps({"status": "success", key: value})
 
-    def _set_tts(self, provider_id: str, model: str) -> str:
+    def _set_tts(self, provider_id: str, model: str, voice: str = "") -> str:
         local_config = conf()
         file_cfg = self._read_file_config()
-        if provider_id:
         local_config["text_to_voice"] = provider_id
         file_cfg["text_to_voice"] = provider_id
-        if model:
         local_config["text_to_voice_model"] = model
         file_cfg["text_to_voice_model"] = model
+        local_config["tts_voice_id"] = voice
+        file_cfg["tts_voice_id"] = voice
         self._write_file_config(file_cfg)
-        logger.info(f"[ModelsHandler] tts updated: provider={provider_id!r} model={model!r}")
+        logger.info(
+            f"[ModelsHandler] tts updated: provider={provider_id!r} "
+            f"model={model!r} voice={voice!r}"
+        )
         self._refresh_voice_routing()
-        return json.dumps({"status": "success", "provider": provider_id, "model": model})
+        return json.dumps({
+            "status": "success",
+            "provider": provider_id, "model": model, "voice": voice,
+        })
 
     @staticmethod
     def _refresh_voice_routing() -> None:
@@ -2066,17 +2529,20 @@ class ModelsHandler:
             logger.warning(f"[ModelsHandler] Bridge voice refresh failed: {e}")
 
     def _set_embedding(self, provider_id: str, model: str) -> str:
-        # provider_id="" + model="" means "switch back to legacy auto mode".
+        # Two valid states: both empty (reset to pick-or-empty) OR both set.
+        # A provider without a model leaves the runtime in a broken half-state,
+        # so reject that explicitly instead of silently writing it through.
+        if provider_id and not model:
+            return json.dumps({
+                "status": "error",
+                "message": "embedding model is required when a provider is selected",
+            })
         local_config = conf()
         file_cfg = self._read_file_config()
         local_config["embedding_provider"] = provider_id
         file_cfg["embedding_provider"] = provider_id
-        if model:
         local_config["embedding_model"] = model
         file_cfg["embedding_model"] = model
-        else:
-            local_config["embedding_model"] = ""
-            file_cfg["embedding_model"] = ""
         self._write_file_config(file_cfg)
         logger.info(f"[ModelsHandler] embedding updated: provider={provider_id!r} model={model!r}")
         # The next /memory rebuild-index command hot-swaps the provider onto

config-template.json

@@ -16,8 +16,8 @@
   "open_ai_api_base": "https://api.openai.com/v1",
   "gemini_api_key": "",
   "gemini_api_base": "https://generativelanguage.googleapis.com",
-  "voice_to_text": "openai",
-  "text_to_voice": "openai",
+  "voice_to_text": "",
+  "text_to_voice": "",
   "voice_reply_voice": false,
   "speech_recognition": true,
   "group_speech_recognition": false,

config.py

@@ -330,8 +330,18 @@ def load_config():
     config_str = read_file(config_path)
     logger.debug("[INIT] config str: {}".format(drag_sensitive(config_str)))
 
-    # 将json字符串反序列化为dict类型
-    config = Config(json.loads(config_str))
+    # 将json字符串反序列化为dict类型。
+    # `object_pairs_hook` lets us catch users who accidentally typed the
+    # same key twice (e.g. two `"tools"` blocks) — json.loads would
+    # otherwise silently drop all but the last occurrence.
+    config = Config(json.loads(config_str, object_pairs_hook=_merge_duplicate_keys))
+
+    # Migrate legacy singular keys (`tool`, `skill`) into the canonical
+    # plural buckets so the rest of the codebase only reads one schema.
+    # Deep-merge so existing `tools`/`skills` entries are preserved and
+    # only missing namespaces are filled in from the legacy section.
+    _merge_legacy_namespace(config, legacy="tool",  canonical="tools")
+    _merge_legacy_namespace(config, legacy="skill", canonical="skills")
 
     # override config with environment variables.
     # Some online deployment platforms (e.g. Railway) deploy project from github directly. So you shouldn't put your secrets like api key in a config file, instead use environment variables to override the default config.
@@ -422,7 +432,7 @@ def load_config():
                 os.environ[env_key] = str(val)
                 injected += 1
 
-    injected += _sync_skill_config_to_env(config.get("skill", {}))
+    injected += _sync_skill_config_to_env(config.get("skills", {}))
 
     if injected:
         logger.info("[INIT] Synced {} config values to environment variables".format(injected))
@@ -430,11 +440,90 @@ def load_config():
     config.load_user_datas()
 
 
+def _deep_merge_dicts(base: dict, incoming: dict) -> dict:
+    """Recursively merge ``incoming`` into ``base`` (incoming wins on leaves)."""
+    for key, val in incoming.items():
+        if (
+            key in base
+            and isinstance(base[key], dict)
+            and isinstance(val, dict)
+        ):
+            _deep_merge_dicts(base[key], val)
+        else:
+            base[key] = val
+    return base
+
+
+def _merge_duplicate_keys(pairs):
+    """object_pairs_hook for json.loads: deep-merge duplicate top-level keys
+    (lists concat, dicts merge, scalars take the latter) instead of dropping."""
+    out = {}
+    duplicates = []
+    for key, val in pairs:
+        if key not in out:
+            out[key] = val
+            continue
+        duplicates.append(key)
+        prev = out[key]
+        if isinstance(prev, dict) and isinstance(val, dict):
+            _deep_merge_dicts(prev, val)
+        elif isinstance(prev, list) and isinstance(val, list):
+            prev.extend(val)
+        else:
+            out[key] = val
+    if duplicates:
+        # logger may not be wired yet — fall back to print so we never lose the warning.
+        unique = sorted(set(duplicates))
+        try:
+            logger.warning("[INIT] config.json has duplicate keys (merged): %s", unique)
+        except Exception:
+            print("[INIT] config.json has duplicate keys (merged):", unique)
+    return out
+
+
+def _merge_legacy_namespace(cfg, legacy: str, canonical: str) -> None:
+    """Fold deprecated singular keys (``tool`` / ``skill``) into their plural
+    canonical counterparts at load time. Canonical entries always win."""
+    legacy_section = cfg.get(legacy)
+    if not isinstance(legacy_section, dict) or not legacy_section:
+        cfg.pop(legacy, None)
+        return
+    canonical_section = cfg.get(canonical)
+    if not isinstance(canonical_section, dict):
+        canonical_section = {}
+    merged_keys = []
+    for name, val in legacy_section.items():
+        if name in canonical_section:
+            if isinstance(canonical_section[name], dict) and isinstance(val, dict):
+                for sub_key, sub_val in val.items():
+                    if (
+                        sub_key in canonical_section[name]
+                        and isinstance(canonical_section[name][sub_key], dict)
+                        and isinstance(sub_val, dict)
+                    ):
+                        _deep_merge_dicts(sub_val, canonical_section[name][sub_key])
+                        canonical_section[name][sub_key] = sub_val
+                    else:
+                        canonical_section[name].setdefault(sub_key, sub_val)
+            continue
+        canonical_section[name] = val
+        merged_keys.append(name)
+    cfg[canonical] = canonical_section
+    cfg.pop(legacy, None)
+    if merged_keys:
+        logger.warning(
+            "[INIT] Legacy config key '{}' is deprecated; merged into '{}': {}. "
+            "Please rename '{}' to '{}' in your config.json.".format(
+                legacy, canonical, merged_keys, legacy, canonical,
+            )
+        )
+
+
 def _sync_skill_config_to_env(skill_section) -> int:
     """Flatten skill-namespaced config into environment variables.
 
-    Mapping rule: ``config["skill"][<name>][<key>]`` -> ``SKILL_<NAME>_<KEY>``
-    (e.g. ``skill["image-generation"].model`` -> ``SKILL_IMAGE_GENERATION_MODEL``).
+    Mapping rule: ``config["skills"][<name>][<key>]`` -> ``SKILL_<NAME>_<KEY>``
+    (e.g. ``skills["image-generation"].model`` -> ``SKILL_IMAGE_GENERATION_MODEL``).
 
     This lets subprocess-based skill scripts read their own settings without
     importing project code. Existing env vars are NOT overwritten so the

docs/en/models/qianfan.mdx

@@ -40,7 +40,7 @@ To force a specific Vision model, set it explicitly in `config.json`:
 
 ```json
 {
-  "tool": {
+  "tools": {
     "vision": {
       "model": "ernie-4.5-turbo-vl"
     }

docs/en/releases/v2.0.7.mdx

@@ -11,7 +11,7 @@ New built-in `image-generation` skill supporting text-to-image, image-to-image,
 - **Zero model selection**: Just configure an API key and it works — no need to manually specify a model. You can also name a specific model in conversation (e.g. "draw a cat with seedream")
 - **Flexible control**: Supports `quality`, `size` (512/1K–4K), and `aspect_ratio` parameters, with each provider automatically mapping to its supported values
 - **Image editing**: Pass existing images for editing, style transfer, or multi-image fusion (Seedream supports up to 14 reference images)
-- **Skill-level config**: Pin a default model via `skill.image-generation.model` in `config.json`
+- **Skill-level config**: Pin a default model via `skills.image-generation.model` in `config.json`
 - **Image lightbox**: All images in the Web console now support click-to-enlarge preview
 
 Docs: [Image Generation Skill](https://docs.cowagent.ai/en/skills/image-generation)

docs/en/releases/v2.0.8.mdx

@@ -51,7 +51,7 @@ The voice and streaming building blocks come from a community contribution #2791
 
 ## 🔧 Tools and Safety
 
-- **Vision model selection**: `tool.vision.model` config now actually takes effect, with automatic fallback when unconfigured #2792
+- **Vision model selection**: `tools.vision.model` config now actually takes effect, with automatic fallback when unconfigured #2792
 - **Bash safety prompt**: The destructive-deletion confirm prompt is now scoped to paths outside the workspace — routine in-workspace operations are no longer interrupted
 
 ## 🐛 Other Fixes

docs/en/skills/image-generation.mdx

@@ -87,7 +87,7 @@ Configure ARK_API_KEY as xxx
 To force all image generation through a specific provider's model, add this to `config.json`:
 
 ```json
-"skill": {
+"skills": {
   "image-generation": {
     "model": "seedream-5.0-lite"
   }

docs/en/tools/vision.mdx

@@ -51,7 +51,7 @@ To specify a particular model for the vision tool, add to `config.json`:
 
 ```json
 {
-    "tool": {
+    "tools": {
         "vision": {
             "model": "ernie-4.5-turbo-vl"
         }

docs/ja/models/qianfan.mdx

@@ -40,7 +40,7 @@ description: Baidu Qianfan ERNIE モデル設定
 
 ```json
 {
-  "tool": {
+  "tools": {
     "vision": {
       "model": "ernie-4.5-turbo-vl"
     }

docs/ja/releases/v2.0.7.mdx

@@ -11,7 +11,7 @@ description: CowAgent 2.0.7 - 画像生成スキル（6プロバイダー自動
 - **モデル選択不要**：API Key を設定するだけで使用可能、モデルを手動で指定する必要なし。会話で特定モデルを指名することも可能（例：「seedream で猫を描いて」）
 - **柔軟な制御**：`quality`（画質）、`size`（解像度、512/1K〜4K）、`aspect_ratio`（アスペクト比）パラメータ対応、各プロバイダーが自動的に有効な値にマッピング
 - **画像編集**：既存の画像を渡して編集・スタイル変換・複数画像融合が可能（Seedream は最大 14 枚の参照画像をサポート）
-- **スキルレベル設定**：`config.json` の `skill.image-generation.model` でデフォルトモデルを固定可能
+- **スキルレベル設定**：`config.json` の `skills.image-generation.model` でデフォルトモデルを固定可能
 - **画像ライトボックス**：Web コンソールのすべての画像がクリックで拡大プレビュー対応
 
 ドキュメント：[画像生成スキル](https://docs.cowagent.ai/ja/skills/image-generation)

docs/ja/releases/v2.0.8.mdx

@@ -51,7 +51,7 @@ description: CowAgent 2.0.8 - 飛書チャネル全面アップグレード（
 
 ## 🔧 ツールと安全性
 
-- **Vision モデル選択**：`tool.vision.model` 設定が実際に反映されるようになり、未設定時は自動フォールバック #2792
+- **Vision モデル選択**：`tools.vision.model` 設定が実際に反映されるようになり、未設定時は自動フォールバック #2792
 - **Bash セーフティ確認**：破壊的削除の確認プロンプトをワークスペース外のパスに限定。ワークスペース内の通常操作は中断されません
 
 ## 🐛 その他の修正

docs/ja/skills/image-generation.mdx

@@ -87,7 +87,7 @@ ARK_API_KEY を xxx に設定して
 すべての画像生成を特定のプロバイダーのモデルで固定したい場合、`config.json` に以下を追加：
 
 ```json
-"skill": {
+"skills": {
   "image-generation": {
     "model": "seedream-5.0-lite"
   }

docs/ja/tools/vision.mdx

@@ -51,7 +51,7 @@ Vision ツールで使用するモデルを指定するには、`config.json`
 
 ```json
 {
-    "tool": {
+    "tools": {
         "vision": {
             "model": "ernie-4.5-turbo-vl"
         }

docs/models/qianfan.mdx

@@ -40,7 +40,7 @@ description: 百度千帆 ERNIE 模型配置
 
 ```json
 {
-  "tool": {
+  "tools": {
     "vision": {
       "model": "ernie-4.5-turbo-vl"
     }

docs/releases/v2.0.7.mdx

@@ -11,7 +11,7 @@ description: CowAgent 2.0.7 - 图像生成技能（六厂商自动路由）、
 - **开箱即用**：配置 API Key 即可使用，无需手动指定模型。也支持在对话中指定特定模型
 - **灵活控制**：支持 `quality`（画质）、`size`（分辨率，512/1K~4K）、`aspect_ratio`（宽高比）等参数，各厂商自动适配有效值
 - **图片编辑**：传入已有图片即可进行编辑、风格迁移、多图融合
-- **Skill 级配置**：支持通过 `config.json` 中的 `skill.image-generation.model` 固定默认模型
+- **Skill 级配置**：支持通过 `config.json` 中的 `skills.image-generation.model` 固定默认模型
 
 相关文档：[图像生成技能](https://docs.cowagent.ai/skills/image-generation)
 

docs/releases/v2.0.8.mdx

@@ -46,7 +46,7 @@ description: CowAgent 2.0.8 - 飞书渠道全面升级（语音、流式打字
 
 ## 🔧 工具与安全
 
-- **图像识别模型**：让 `tool.vision.model` 配置真正生效，未配置时自动 fallback #2792 Thanks CNXudiandian
+- **图像识别模型**：让 `tools.vision.model` 配置真正生效，未配置时自动 fallback #2792 Thanks CNXudiandian
 - **Bash 安全确认**：仅对工作区外的破坏性删除做二次确认，工作区内常规操作不再打扰
 
 ## 🐛 其他修复

docs/skills/image-generation.mdx

@@ -88,7 +88,7 @@ description: 文生图 / 图生图 / 多图融合，支持多家厂商自动路
 如果想让所有图像生成固定走某个厂商的模型，可以在 `config.json` 里加：
 
 ```json
-"skill": {
+"skills": {
   "image-generation": {
     "model": "seedream-5.0-lite"
   }

docs/tools/vision.mdx

@@ -40,7 +40,7 @@ Vision 工具采用多级自动选择 + 自动兜底策略，无需手动配置
 
 ```json
 {
-    "tool": {
+    "tools": {
         "vision": {
             "model": "gpt-4.1"
         }

skills/image-generation/scripts/generate.py

@@ -1110,7 +1110,7 @@ def main():
     # Model resolution priority:
     #   1. Explicit `model` in the call args (agent / user override)
     #   2. SKILL_IMAGE_GENERATION_MODEL env var (synced from
-    #      config["skill"]["image-generation"]["model"] at startup)
+    #      config["skills"]["image-generation"]["model"] at startup)
     #   3. None → fall back to automatic provider routing (try every
     #      provider with a configured API key in global priority order)
     model = args.get("model") or os.environ.get("SKILL_IMAGE_GENERATION_MODEL") or ""

tests/test_qianfan_provider.py

@@ -394,7 +394,7 @@ class TestQianfanVisionTool(unittest.TestCase):
             "open_ai_api_key": "",
             "linkai_api_key": "",
             "use_linkai": False,
-            "tool": {},
+            "tools": {},
         }
         if values:
             data.update(values)
@@ -424,7 +424,7 @@ class TestQianfanVisionTool(unittest.TestCase):
     def test_vision_routes_ernie_model_override_to_qianfan(self):
         fake_conf = self._fake_conf({
             "qianfan_api_key": "test-qianfan-key",
-            "tool": {"vision": {"model": "ernie-4.5-turbo-vl-32k"}},
+            "tools": {"vision": {"model": "ernie-4.5-turbo-vl-32k"}},
         })
         fake_bot = MagicMock()
         fake_bot.call_vision = MagicMock()

voice/dashscope/dashscope_voice.py

@@ -1,20 +1,13 @@
 # encoding:utf-8
-"""
-DashScope (Aliyun Bailian) voice service.
-
-ASR : qwen3-asr-flash via dashscope.MultiModalConversation
-TTS : not yet implemented (see CosyVoice / qwen3-tts)
-
-Why MultiModalConversation instead of the OpenAI-compatible endpoint:
-  - SDK is already a project dep (used by chat/vision)
-  - Native API accepts local file:// paths up to 100 QPS without an OSS
-    round-trip, which is what we need for the "send a short voice
-    message" flow. Public URLs / Base64 also work.
-"""
+"""DashScope voice: qwen3-asr-flash (ASR) + qwen3-tts-flash (TTS)
+via dashscope.MultiModalConversation."""
+import datetime
 import os
+import random
 from typing import Optional
 
 import dashscope
+import requests
 from dashscope import MultiModalConversation
 
 from bridge.reply import Reply, ReplyType
@@ -25,16 +18,14 @@ from voice.voice import Voice
 
 
 DEFAULT_ASR_MODEL = "qwen3-asr-flash"
-# qwen3-asr-flash hard cap (single file, sync call). Longer audio needs
-# qwen3-asr-flash-filetrans which is async-only and out of scope here.
+DEFAULT_TTS_MODEL = "qwen3-tts-flash"
+DEFAULT_TTS_VOICE = "Cherry"
 MAX_DURATION_SECONDS = 300
 MAX_FILE_BYTES = 10 * 1024 * 1024
 
 
 class DashScopeVoice(Voice):
     def __init__(self):
-        # api_key is applied per-call (chat bot does the same) so a live
-        # config change via the web console takes effect without restart.
         pass
 
     def voiceToText(self, voice_file: str):
@@ -83,14 +74,72 @@ class DashScopeVoice(Voice):
             return Reply(ReplyType.ERROR, "我暂时还无法听清您的语音，请稍后再试吧~")
 
     def textToVoice(self, text: str):
-        # TTS will be added in a follow-up commit (qwen3-tts / cosyvoice).
-        return Reply(ReplyType.ERROR, "DashScope 语音合成尚未接入")
+        try:
+            api_key = conf().get("dashscope_api_key", "")
+            if not api_key:
+                logger.error("[DashScopeVoice] dashscope_api_key is not configured")
+                return Reply(ReplyType.ERROR, "未配置 DashScope API key")
+            dashscope.api_key = api_key
+
+            model = conf().get("text_to_voice_model") or DEFAULT_TTS_MODEL
+            voice = conf().get("tts_voice_id") or DEFAULT_TTS_VOICE
+            response = MultiModalConversation.call(
+                model=model,
+                api_key=api_key,
+                text=text,
+                voice=voice,
+                stream=False,
+            )
+
+            url = self._extract_audio_url(response)
+            if not url:
+                logger.error(f"[DashScopeVoice] textToVoice failed: {response}")
+                return Reply(ReplyType.ERROR, "语音合成失败")
+
+            local_path = self._download_audio(url)
+            if not local_path:
+                return Reply(ReplyType.ERROR, "语音合成失败")
+
+            logger.info(f"[DashScopeVoice] textToVoice model={model} voice={voice} file={local_path}")
+            return Reply(ReplyType.VOICE, local_path)
+        except Exception as e:
+            logger.exception(f"[DashScopeVoice] textToVoice exception: {e}")
+            return Reply(ReplyType.ERROR, "语音合成失败")
+
+    @staticmethod
+    def _extract_audio_url(response) -> Optional[str]:
+        try:
+            if getattr(response, "status_code", 200) != 200:
+                return None
+            audio = response.output.get("audio") if response.output else None
+            if isinstance(audio, dict):
+                return audio.get("url") or None
+            return getattr(audio, "url", None)
+        except Exception:
+            return None
+
+    @staticmethod
+    def _download_audio(url: str) -> Optional[str]:
+        try:
+            tmp_dir = os.path.join(os.getcwd(), "tmp")
+            os.makedirs(tmp_dir, exist_ok=True)
+            ts = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
+            ext = os.path.splitext(url.split("?", 1)[0])[1].lower() or ".wav"
+            if ext not in (".mp3", ".wav", ".m4a", ".aac", ".opus"):
+                ext = ".wav"
+            dst = os.path.join(tmp_dir, f"dashscope_tts_{ts}_{random.randint(0, 9999)}{ext}")
+            resp = requests.get(url, timeout=60)
+            resp.raise_for_status()
+            with open(dst, "wb") as f:
+                f.write(resp.content)
+            return dst
+        except Exception as e:
+            logger.error(f"[DashScopeVoice] download audio failed: {e}")
+            return None
 
     @staticmethod
     def _ensure_compatible_format(voice_file: str) -> str:
-        """Convert AMR/SILK to mp3 since qwen3-asr-flash doesn't accept them.
-        Other formats (mp3/wav/m4a/aac/opus/webm) are passed through.
-        """
+        # qwen3-asr-flash doesn't accept AMR/SILK; mp3/wav/m4a/aac/opus pass through.
         lower = voice_file.lower()
         if lower.endswith(".amr") or lower.endswith(".silk") or lower.endswith(".slk"):
             try:
@@ -98,20 +147,11 @@ class DashScopeVoice(Voice):
                 audio_convert.any_to_mp3(voice_file, mp3_file)
                 return mp3_file
             except Exception as e:
-                logger.warning(
-                    f"[DashScopeVoice] convert {voice_file} to mp3 failed: {e}; "
-                    f"submitting original file"
-                )
+                logger.warning(f"[DashScopeVoice] mp3 convert failed: {e}")
         return voice_file
 
     @staticmethod
     def _extract_text(response) -> Optional[str]:
-        """Pull the recognized text out of MultiModalConversation response.
-
-        Successful shape (result_format="message"):
-          response.output.choices[0].message.content -> list of {"text": "..."}
-          or in some SDK versions a plain string.
-        """
         try:
             if getattr(response, "status_code", 200) != 200:
                 return None

voice/linkai/linkai_voice.py

@@ -1,16 +1,18 @@
-"""
-google voice service
-"""
+"""LinkAI voice: Whisper ASR + multi-vendor TTS (OpenAI / Doubao / Baidu)
+proxied via https://docs.link-ai.tech/platform/api/voice-speech."""
+import datetime
+import os
 import random
+
 import requests
-from voice import audio_convert
+
 from bridge.reply import Reply, ReplyType
+from common import const
 from common.log import logger
 from config import conf
+from voice import audio_convert
 from voice.voice import Voice
-from common import const
-import os
-import datetime
+
 
 class LinkAIVoice(Voice):
     def __init__(self):
@@ -21,8 +23,7 @@ class LinkAIVoice(Voice):
         try:
             url = conf().get("linkai_api_base", "https://api.link-ai.tech") + "/v1/audio/transcriptions"
             headers = {"Authorization": "Bearer " + conf().get("linkai_api_key")}
-            model = None
-            if not conf().get("text_to_voice") or conf().get("voice_to_text") == "openai":
+            # Pin whisper-1: gateway ignores any other ASR model id.
             model = const.WHISPER_1
             if voice_file.endswith(".amr"):
                 try:
@@ -30,54 +31,59 @@ class LinkAIVoice(Voice):
                     audio_convert.any_to_mp3(voice_file, mp3_file)
                     voice_file = mp3_file
                 except Exception as e:
-                    logger.warn(f"[LinkVoice] amr file transfer failed, directly send amr voice file: {format(e)}")
-            file = open(voice_file, "rb")
-            file_body = {
-                "file": file
-            }
-            data = {
-                "model": model
-            }
-            res = requests.post(url, files=file_body, headers=headers, data=data, timeout=(5, 60))
-            if res.status_code == 200:
-                text = res.json().get("text")
-            else:
-                res_json = res.json()
-                logger.error(f"[LinkVoice] voiceToText error, status_code={res.status_code}, msg={res_json.get('message')}")
+                    logger.warning(f"[LinkVoice] amr file transfer failed, directly send amr voice file: {e}")
+            with open(voice_file, "rb") as file:
+                res = requests.post(
+                    url,
+                    files={"file": file},
+                    headers=headers,
+                    data={"model": model},
+                    timeout=(5, 60),
+                )
+            if res.status_code != 200:
+                msg = ""
+                try:
+                    msg = res.json().get("message", "")
+                except Exception:
+                    pass
+                logger.error(f"[LinkVoice] voiceToText error, status_code={res.status_code}, msg={msg}")
                 return None
-            reply = Reply(ReplyType.TEXT, text)
+            text = res.json().get("text")
             logger.info(f"[LinkVoice] voiceToText success, text={text}, file name={voice_file}")
+            return Reply(ReplyType.TEXT, text)
         except Exception as e:
             logger.error(e)
             return None
-        return reply
 
     def textToVoice(self, text):
         try:
             url = conf().get("linkai_api_base", "https://api.link-ai.tech") + "/v1/audio/speech"
             headers = {"Authorization": "Bearer " + conf().get("linkai_api_key")}
-            model = const.TTS_1
-            if not conf().get("text_to_voice") or conf().get("text_to_voice") in ["openai", const.TTS_1, const.TTS_1_HD]:
-                model = conf().get("text_to_voice_model") or const.TTS_1
+            # Gateway routes by `model` (tts-1 / doubao / baidu) + `voice` from
+            # that engine's catalog. `app_code` is optional workspace override.
             data = {
-                "model": model,
                 "input": text,
                 "voice": conf().get("tts_voice_id"),
-                "app_code": conf().get("linkai_app_code")
+                "app_code": conf().get("linkai_app_code"),
             }
+            model = conf().get("text_to_voice_model")
+            if model:
+                data["model"] = model
             res = requests.post(url, headers=headers, json=data, timeout=(5, 120))
-            if res.status_code == 200:
+            if res.status_code != 200:
+                msg = ""
+                try:
+                    msg = res.json().get("message", "")
+                except Exception:
+                    pass
+                logger.error(f"[LinkVoice] textToVoice error, status_code={res.status_code}, msg={msg}")
+                return None
             tmp_file_name = "tmp/" + datetime.datetime.now().strftime('%Y%m%d%H%M%S') + str(random.randint(0, 1000)) + ".mp3"
+            os.makedirs(os.path.dirname(tmp_file_name), exist_ok=True)
             with open(tmp_file_name, 'wb') as f:
                 f.write(res.content)
-                reply = Reply(ReplyType.VOICE, tmp_file_name)
-                logger.info(f"[LinkVoice] textToVoice success, input={text}, model={model}, voice_id={data.get('voice')}")
-                return reply
-            else:
-                res_json = res.json()
-                logger.error(f"[LinkVoice] textToVoice error, status_code={res.status_code}, msg={res_json.get('message')}")
-                return None
+            logger.info(f"[LinkVoice] textToVoice success, input={text}, voice_id={data.get('voice')}")
+            return Reply(ReplyType.VOICE, tmp_file_name)
         except Exception as e:
             logger.error(e)
-            # reply = Reply(ReplyType.ERROR, "遇到了一点小问题，请稍后再问我吧")
             return None

voice/minimax/minimax_voice.py

@@ -1,8 +1,7 @@
 # encoding:utf-8
-"""
-MiniMax TTS voice service
-"""
+"""MiniMax TTS via /v1/t2a_v2 (SSE stream, hex-encoded mp3 chunks)."""
 import datetime
+import json
 import random
 import requests
 
@@ -12,24 +11,12 @@ from config import conf
 from voice.voice import Voice
 
 
-MINIMAX_TTS_VOICES = [
-    "English_Graceful_Lady",
-    "English_Insightful_Speaker",
-    "English_radiant_girl",
-    "English_Persuasive_Man",
-    "English_Lucky_Robot",
-    "English_expressive_narrator",
-    "Chinese_Warm_Woman",
-    "Chinese_Gentle_Man",
-]
-
-
 class MinimaxVoice(Voice):
     def __init__(self):
         self.api_key = conf().get("minimax_api_key")
-        self.api_base = conf().get("minimax_api_base") or "https://api.minimax.io"
-        # Strip trailing /v1 if present so we can always append /v1/t2a_v2
-        self.api_base = self.api_base.rstrip("/")
+        # Mainland endpoint matches `sk-api-0-...` keys; override via
+        # `minimax_api_base` for international (api.minimax.io) workspaces.
+        self.api_base = (conf().get("minimax_api_base") or "https://api.minimaxi.com").rstrip("/")
         if self.api_base.endswith("/v1"):
             self.api_base = self.api_base[:-3]
 
@@ -68,12 +55,14 @@ class MinimaxVoice(Voice):
             response = requests.post(url, headers=headers, json=payload, stream=True, timeout=60)
             response.raise_for_status()
 
-            # Parse SSE stream and collect hex-encoded audio chunks
+            # MiniMax returns HTTP 200 even on errors; capture base_resp for diagnostics.
             audio_chunks = []
-            buffer = ""
+            last_base_resp = None
+            event_count = 0
             for raw in response.iter_lines():
                 if not raw:
                     continue
+                event_count += 1
                 line = raw.decode("utf-8") if isinstance(raw, bytes) else raw
                 if not line.startswith("data:"):
                     continue
@@ -81,16 +70,31 @@ class MinimaxVoice(Voice):
                 if not json_str or json_str == "[DONE]":
                     continue
                 try:
-                    import json
                     event_data = json.loads(json_str)
-                    audio_hex = event_data.get("data", {}).get("audio")
-                    if audio_hex:
-                        audio_chunks.append(bytes.fromhex(audio_hex))
                 except Exception:
                     continue
+                base_resp = event_data.get("base_resp") or {}
+                if base_resp:
+                    last_base_resp = base_resp
+                audio_hex = (event_data.get("data") or {}).get("audio")
+                if audio_hex:
+                    try:
+                        audio_chunks.append(bytes.fromhex(audio_hex))
+                    except Exception as e:
+                        logger.warning(f"[MINIMAX] skip bad audio hex chunk: {e}")
 
             if not audio_chunks:
-                logger.error("[MINIMAX] TTS returned no audio data")
+                ct = response.headers.get("Content-Type", "")
+                if last_base_resp and last_base_resp.get("status_code") not in (None, 0):
+                    logger.error(
+                        f"[MINIMAX] TTS failed: status_code={last_base_resp.get('status_code')}, "
+                        f"status_msg={last_base_resp.get('status_msg')}, model={model}, voice_id={voice_id}"
+                    )
+                else:
+                    logger.error(
+                        f"[MINIMAX] TTS returned no audio data, model={model}, voice_id={voice_id}, "
+                        f"url={url}, http={response.status_code}, content_type={ct!r}, events={event_count}"
+                    )
                 return Reply(ReplyType.ERROR, "语音合成失败，未获取到音频数据")
 
             audio_data = b"".join(audio_chunks)

voice/openai/openai_voice.py

@@ -31,7 +31,8 @@ class OpenaiVoice(Voice):
                 "file": file,
             }
             data = {
-                "model": "whisper-1",
+                # Override via `voice_to_text_model` (e.g. fall back to whisper-1).
+                "model": conf().get("voice_to_text_model") or "gpt-4o-mini-transcribe",
             }
             response = requests.post(url, headers=headers, files=files, data=data)
             response_data = response.json()

voice/zhipuai/zhipuai_voice.py

@@ -1,14 +1,8 @@
 # encoding:utf-8
-"""
-ZhipuAI (BigModel) voice service.
-
-ASR : glm-asr-2512 via the OpenAI-compatible /audio/transcriptions endpoint.
-TTS : not yet implemented.
-
-Endpoint accepts multipart/form-data with `model`, `file`, and `stream`.
-File size <= 25MB, duration <= 30s per request.
-"""
+"""ZhipuAI voice: glm-asr-2512 (ASR) + glm-tts (TTS) via BigModel REST API."""
+import datetime
 import os
+import random
 
 import requests
 
@@ -20,6 +14,8 @@ from voice.voice import Voice
 
 
 DEFAULT_ASR_MODEL = "glm-asr-2512"
+DEFAULT_TTS_MODEL = "glm-tts"
+DEFAULT_TTS_VOICE = "tongtong"
 DEFAULT_API_BASE = "https://open.bigmodel.cn/api/paas/v4"
 MAX_FILE_BYTES = 25 * 1024 * 1024
 REQUEST_TIMEOUT = (5, 60)
@@ -27,7 +23,6 @@ REQUEST_TIMEOUT = (5, 60)
 
 class ZhipuAIVoice(Voice):
     def __init__(self):
-        # api_key / base read per-call so live config edits take effect.
         pass
 
     def voiceToText(self, voice_file: str):
@@ -81,12 +76,91 @@ class ZhipuAIVoice(Voice):
             return Reply(ReplyType.ERROR, "我暂时还无法听清您的语音，请稍后再试吧~")
 
     def textToVoice(self, text: str):
-        return Reply(ReplyType.ERROR, "ZhipuAI 语音合成尚未接入")
+        try:
+            api_key = conf().get("zhipu_ai_api_key", "")
+            if not api_key:
+                logger.error("[ZhipuAIVoice] zhipu_ai_api_key is not configured")
+                return Reply(ReplyType.ERROR, "未配置 ZhipuAI API key")
+
+            api_base = (conf().get("zhipu_ai_api_base") or DEFAULT_API_BASE).rstrip("/")
+            url = f"{api_base}/audio/speech"
+            model = conf().get("text_to_voice_model") or DEFAULT_TTS_MODEL
+            voice_id = conf().get("tts_voice_id") or DEFAULT_TTS_VOICE
+
+            payload = {
+                "model": model,
+                "input": text,
+                "voice": voice_id,
+                "response_format": "wav",
+                "speed": 1.0,
+                "volume": 1.0,
+            }
+            headers = {
+                "Authorization": f"Bearer {api_key}",
+                "Content-Type": "application/json",
+            }
+            response = requests.post(
+                url, headers=headers, json=payload, timeout=REQUEST_TIMEOUT
+            )
+
+            if response.status_code != 200:
+                logger.error(
+                    f"[ZhipuAIVoice] textToVoice failed: status={response.status_code} "
+                    f"body={response.text[:500]} model={model} voice={voice_id}"
+                )
+                return Reply(ReplyType.ERROR, "语音合成失败，请稍后再试")
+
+            # Some errors come back as JSON / SSE with HTTP 200.
+            ct = response.headers.get("Content-Type", "")
+            if "application/json" in ct or "text/event-stream" in ct:
+                try:
+                    err = response.json()
+                except Exception:
+                    err = {"raw": response.text[:500]}
+                logger.error(
+                    f"[ZhipuAIVoice] textToVoice unexpected text response "
+                    f"(content_type={ct}): {err}"
+                )
+                return Reply(ReplyType.ERROR, "语音合成失败，请稍后再试")
+
+            audio_bytes = response.content
+            ext = self._sniff_audio_ext(audio_bytes) or "wav"
+
+            file_name = (
+                "tmp/" + datetime.datetime.now().strftime("%Y%m%d%H%M%S")
+                + str(random.randint(0, 1000)) + "." + ext
+            )
+            os.makedirs(os.path.dirname(file_name), exist_ok=True)
+            with open(file_name, "wb") as f:
+                f.write(audio_bytes)
+            logger.info(
+                f"[ZhipuAIVoice] textToVoice model={model} voice={voice_id} "
+                f"file={file_name} bytes={len(audio_bytes)} ext={ext}"
+            )
+            return Reply(ReplyType.VOICE, file_name)
+        except Exception as e:
+            logger.exception(f"[ZhipuAIVoice] textToVoice exception: {e}")
+            return Reply(ReplyType.ERROR, "语音合成失败，请稍后再试")
+
+    @staticmethod
+    def _sniff_audio_ext(data: bytes) -> str:
+        """Detect audio container by magic bytes; returns '' on unknown."""
+        if len(data) < 12:
+            return ""
+        head = data[:12]
+        if head[:4] == b"RIFF" and head[8:12] == b"WAVE":
+            return "wav"
+        if head[:3] == b"ID3" or head[:2] == b"\xff\xfb" or head[:2] == b"\xff\xf3" or head[:2] == b"\xff\xf2":
+            return "mp3"
+        if head[:4] == b"OggS":
+            return "ogg"
+        if head[:4] == b"fLaC":
+            return "flac"
+        return ""
 
     @staticmethod
     def _ensure_compatible_format(voice_file: str) -> str:
-        # glm-asr-2512 only accepts .wav / .mp3 — convert everything else
-        # (webm from the browser mic, m4a/amr/silk from chat channels, etc).
+        # glm-asr-2512 only accepts .wav / .mp3
         lower = voice_file.lower()
         if lower.endswith(".mp3") or lower.endswith(".wav"):
             return voice_file
@@ -95,8 +169,5 @@ class ZhipuAIVoice(Voice):
             audio_convert.any_to_mp3(voice_file, mp3_file)
             return mp3_file
         except Exception as e:
-            logger.warning(
-                f"[ZhipuAIVoice] convert {voice_file} to mp3 failed: {e}; "
-                f"submitting original file"
-            )
+            logger.warning(f"[ZhipuAIVoice] mp3 convert failed: {e}")
             return voice_file