@ducci/jarvis 1.0.92 → 1.0.94

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ducci/jarvis",
-  "version": "1.0.92",
+  "version": "1.0.94",
   "description": "A fully automated agent system that lives on a server.",
   "main": "./src/index.js",
   "type": "module",

package/src/channels/telegram/index.js

@@ -697,7 +697,6 @@ export async function startTelegramChannel(config) {
           await sendMessage(api, chatId, displayText, result.sessionId);
           console.log(`[telegram] response sent chat_id=${chatId} slot=${slot} length=${displayText.length}`);
           // TTS: send audio summary if voice is enabled (config.voiceEnabled checked live, updated by /voice toggle)
-          await api.sendMessage(chatId, `[TTS debug] voiceEnabled=${config.voiceEnabled} hasKey=${!!config.fishAudioApiKey}`).catch(() => {});
           if (config.voiceEnabled && config.fishAudioApiKey) {
             try {
               // If the response is a raw JSON blob (format_error recovery), extract the actual text
@@ -709,10 +708,8 @@ export async function startTelegramChannel(config) {
               const plain = toPlainText(ttsSource);
               if (plain) {
                 const ttsText = await generateTtsSummary(plain, config);
-                await api.sendMessage(chatId, `[TTS debug] plain: "${plain.slice(0, 300)}"\nsummary: "${ttsText}"`).catch(() => {});
                 if (ttsText) {
                   const audioBuffer = await textToSpeech(ttsText, config);
-                  await api.sendMessage(chatId, `[TTS debug] audio bytes: ${audioBuffer.length}`).catch(() => {});
                   await api.sendAudio(chatId, new InputFile(audioBuffer, 'response.mp3'));
                   console.log(`[telegram] voice sent chat_id=${chatId} slot=${slot} tts_chars=${ttsText.length}`);
                 }

package/src/server/fish-audio.js

@@ -14,23 +14,7 @@ const execAsync = promisify(exec);
 // fish.audio s1 emotion tags: (emotion) at the START of a sentence only — applies to the whole sentence.
 // Multiple tags can be combined: (excited)(soft tone) Hello!
 // Tags must never appear mid-sentence.
-const TTS_SYSTEM_PROMPT = `You summarize AI assistant responses into a short spoken version for text-to-speech audio.
-
-Rules:
-- Output 1–3 short sentences maximum. Be concise — this is spoken audio, not text.
-- Begin EACH sentence with exactly one emotion tag from this list, placed before the first word:
-  (happy) (sad) (angry) (excited) (calm) (nervous) (confident) (surprised) (satisfied)
-  (delighted) (scared) (worried) (frustrated) (empathetic) (curious) (sarcastic)
-  (optimistic) (determined) (proud) (relaxed) (in a hurry tone) (whispering) (soft tone)
-- The emotion tag goes at the very start of the sentence, before any word. Never mid-sentence.
-- Choose emotions that fit the content: use (confident) or (calm) for informational answers,
-  (excited) or (satisfied) for completed tasks, (curious) for questions, etc.
-- You may combine two tags on one sentence: (excited)(soft tone) Great news!
-- Plain text only — no emojis, no markdown, no code blocks, no bullet points, no special characters.
-- Keep technical jargon minimal; explain concepts simply as you would speak them.
-- Match the language of the original response.
-
-Output only the spoken summary text, nothing else.`;
+const TTS_SYSTEM_PROMPT = `Summarize the given text in 1-2 spoken sentences for audio playback. Start each sentence with an emotion tag like (calm), (excited), (confident), (happy), or (curious) — placed before the first word of the sentence. Plain text only, no emojis, no markdown. Match the language of the text. Output only the summary, nothing else.`;
 
 /**
  * Generate a short TTS-optimized spoken summary of a response via LLM.
@@ -45,14 +29,20 @@ export async function generateTtsSummary(plainText, config) {
 
   const response = await client.chat.completions.create({
     model: config.selectedModel,
-    max_tokens: 200,
     messages: [
       { role: 'system', content: TTS_SYSTEM_PROMPT },
       { role: 'user', content: `Summarize this for spoken audio:\n\n${plainText.slice(0, 3000)}` },
     ],
   });
 
-  return (response.choices[0]?.message?.content || '').trim();
+  const choice = response.choices[0];
+  const msg = choice?.message;
+  const content = (msg?.content || msg?.reasoning_content || '').trim();
+  if (!content) {
+    // Surface the raw response structure so we can diagnose where the text actually is
+    throw new Error(`empty LLM response. finish_reason=${choice?.finish_reason} keys=${Object.keys(msg || {}).join(',')} raw=${JSON.stringify(msg).slice(0, 300)}`);
+  }
+  return content;
 }
 
 /**