neoagent 2.2.1-beta.3 → 2.2.1-beta.5

package/server/routes/settings.js

@@ -48,6 +48,7 @@ const AGENT_SETTING_KEYS = new Set([
   'ai_provider_configs',
   'default_chat_model',
   'default_subagent_model',
+  'default_speech_model',
   'enabled_models',
   'voice_stt_provider',
   'voice_stt_model',

package/server/services/ai/engine.js

@@ -1441,7 +1441,14 @@ class AgentEngine {
   }
 
   async run(userId, userMessage, options = {}) {
-    return this.runWithModel(userId, userMessage, options, null);
+    return this.runWithModel(
+      userId,
+      userMessage,
+      options,
+      typeof options.model === 'string' && options.model.trim()
+        ? options.model.trim()
+        : null,
+    );
   }
 
   async runWithModel(userId, userMessage, options = {}, _modelOverride = null) {

package/server/services/ai/settings.js

@@ -101,6 +101,7 @@ function createDefaultAiSettings() {
     enabled_models: [],
     default_chat_model: 'auto',
     default_subagent_model: 'auto',
+    default_speech_model: 'auto',
     ai_provider_configs: createDefaultProviderConfigs(),
     voice_runtime_mode: 'live',
     voice_live_provider: 'openai',
@@ -284,6 +285,9 @@ function getAiSettings(userId, agentId = null) {
   settings.default_subagent_model = typeof settings.default_subagent_model === 'string' && settings.default_subagent_model.trim()
     ? settings.default_subagent_model
     : DEFAULT_AI_SETTINGS.default_subagent_model;
+  settings.default_speech_model = typeof settings.default_speech_model === 'string' && settings.default_speech_model.trim()
+    ? settings.default_speech_model.trim()
+    : DEFAULT_AI_SETTINGS.default_speech_model;
   settings.voice_runtime_mode = normalizeRuntimeMode(settings.voice_runtime_mode);
   settings.voice_live_provider = normalizeLiveProvider(settings.voice_live_provider);
   settings.voice_live_model = resolveLiveModel(settings.voice_live_provider, settings.voice_live_model);

package/server/services/messaging/automation.js

@@ -154,6 +154,7 @@ async function processQueuedMessage({
     const runOptions = isVoiceLikeMessage(msg)
       ? buildVoiceMessagingRunOptions({
           runId,
+          userId,
           agentId,
           conversationId,
           msg,

package/server/services/voice/providers.js

@@ -595,6 +595,41 @@ async function synthesizeVoiceReply(text, options = {}) {
 
 // Minimum characters before flushing a sentence chunk to TTS to avoid tiny requests.
 const MIN_SENTENCE_CHUNK_CHARS = 80;
+const MAX_TTS_CHUNK_CHARS = 220;
+
+function splitOversizeChunk(text, maxChars = MAX_TTS_CHUNK_CHARS) {
+  const normalized = String(text || '').trim();
+  if (!normalized) return [];
+  if (normalized.length <= maxChars) return [normalized];
+
+  const words = normalized.split(/\s+/).filter(Boolean);
+  if (words.length <= 1) {
+    const slices = [];
+    for (let index = 0; index < normalized.length; index += maxChars) {
+      slices.push(normalized.slice(index, index + maxChars).trim());
+    }
+    return slices.filter(Boolean);
+  }
+
+  const chunks = [];
+  let pending = '';
+  for (const word of words) {
+    const candidate = pending ? `${pending} ${word}` : word;
+    if (pending && candidate.length > maxChars) {
+      chunks.push(pending);
+      pending = word;
+      continue;
+    }
+    if (!pending && candidate.length > maxChars) {
+      chunks.push(...splitOversizeChunk(word, maxChars));
+      pending = '';
+      continue;
+    }
+    pending = candidate;
+  }
+  if (pending) chunks.push(pending);
+  return chunks;
+}
 
 function splitIntoSentenceChunks(text) {
   const normalized = String(text || '').trim();
@@ -609,13 +644,13 @@ function splitIntoSentenceChunks(text) {
     const piece = part.trim();
     if (!piece) continue;
     pending = pending ? `${pending} ${piece}` : piece;
-    if (pending.length >= MIN_SENTENCE_CHUNK_CHARS) {
-      chunks.push(pending);
+    if (pending.length >= MIN_SENTENCE_CHUNK_CHARS || pending.length >= MAX_TTS_CHUNK_CHARS) {
+      chunks.push(...splitOversizeChunk(pending));
       pending = '';
     }
   }
 
-  if (pending) chunks.push(pending);
+  if (pending) chunks.push(...splitOversizeChunk(pending));
   return chunks.length ? chunks : [normalized];
 }
 
@@ -628,8 +663,8 @@ async function synthesizeVoiceReplyStream(text, options = {}, onChunk) {
   const { provider, model, voice } = normalizeVoiceSynthesisOptions(options);
   const chunks = splitIntoSentenceChunks(content);
 
-  const run = (async () => {
-    for (const chunk of chunks) {
+  for (const chunk of chunks) {
+    const run = (async () => {
       if (provider === 'openai') {
         await streamWithOpenAi(chunk, model, voice, options, onChunk);
       } else if (provider === 'deepgram') {
@@ -637,9 +672,9 @@ async function synthesizeVoiceReplyStream(text, options = {}, onChunk) {
       } else {
         await streamWithGemini(chunk, model, voice, options, onChunk);
       }
-    }
-  })();
-  await withTimeout(run, options.timeoutMs, `${provider} TTS stream`);
+    })();
+    await withTimeout(run, options.timeoutMs, `${provider} TTS stream`);
+  }
 }
 
 module.exports = {

package/server/services/voice/runtime.js

@@ -1,6 +1,7 @@
 'use strict';
 
 const { buildPlatformFormattingGuide } = require('../messaging/formatting_guides');
+const { getAiSettings } = require('../ai/settings');
 
 const VOICE_HISTORY_WINDOW = 4;
 const VOICE_REASONING_EFFORT = 'low';
@@ -67,13 +68,17 @@ function buildVoiceMessagingPrompt(msg = {}) {
 
 function buildVoiceMessagingRunOptions({
   runId,
+  userId,
   agentId = null,
   conversationId,
   msg,
 }) {
+  const aiSettings = getAiSettings(userId, agentId);
+  const speechModel = String(aiSettings.default_speech_model || 'auto').trim();
   return {
     runId,
     agentId,
+    model: speechModel !== 'auto' ? speechModel : null,
     triggerSource: 'messaging',
     conversationId,
     source: msg.platform,
@@ -92,12 +97,16 @@ function buildVoiceMessagingRunOptions({
 }
 
 function buildDirectVoiceRunOptions({
+  userId,
   agentId = null,
   conversationId,
   platform = 'voice_assistant',
 }) {
+  const aiSettings = getAiSettings(userId, agentId);
+  const speechModel = String(aiSettings.default_speech_model || 'auto').trim();
   return {
     agentId,
+    model: speechModel !== 'auto' ? speechModel : null,
     conversationId,
     triggerSource: platform,
     skipConversationHistory: true,

package/server/services/voice/runtimeManager.js

@@ -377,7 +377,7 @@ class VoiceRuntimeManager {
         ...normalized,
         apiKey: runtime.apiKey,
         baseUrl: runtime.baseUrl,
-        timeoutMs: 12000,
+        timeoutMs: 20000,
       });
     }
     return attempts;

package/server/services/voice/turnRunner.js

@@ -90,6 +90,7 @@ async function runVoiceTranscriptTurn({
   });
   const conversationId = memoryManager.getDefaultWebConversationId(userId, { agentId });
   const runOptions = buildDirectVoiceRunOptions({
+    userId,
     agentId,
     conversationId,
     platform,