Add gitignor

assistant/cli.py

@@ -1,5 +1,8 @@
 import asyncio
+import queue as _queue
+import re
 import sys
+import threading
 from . import llm, tts, audio, config
 
 
@@ -34,22 +37,122 @@ def _set_voice(parts: list[str]) -> None:
         print(f"Voix définie sur : {config.VOICE_ID}")
 
 
+# ---------------------------------------------------------------------------
+# Sentence splitting & markdown cleaning for TTS pipeline
+# ---------------------------------------------------------------------------
+
+# Split on sentence boundaries using fixed-width lookbehinds (Python constraint).
+# For ".": 3-char lookbehind — require the 2 chars before the period to be
+#   lowercase, digit, hyphen, or closing bracket (NOT uppercase).  This protects
+#   "M.", "No.", "A." while correctly splitting "arrivé.", "Jump-2.", "(J2).", "ok."
+_SENT_RE = re.compile(
+    r'(?<=[a-z\u00e0-\u00ff\d\-)\]]{2}[.])\s+'               # word(2+) + period
+    r'|(?<=[a-z\u00e0-\u00ff\d\-)\]]{2}[.][»"\')\]])\s+'     # + closing quote
+    r'|(?<=[!?…])\s+'                                          # ! ? …
+    r'|(?<=[!?…][»"\')\]])\s+'                                 # ! ? … + closing quote
+    r'|\n{2,}'                                                 # paragraph break
+)
+
+# Markdown patterns to strip before TTS (keep inner text where applicable)
+_MD_CODE_BLOCK = re.compile(r'```.*?```', re.DOTALL)
+_MD_INLINE      = re.compile(r'\*{1,3}(.*?)\*{1,3}|_{1,2}(.*?)_{1,2}|~~(.*?)~~|`([^`]+)`', re.DOTALL)
+_MD_LINK        = re.compile(r'\[([^\]]*)\]\([^\)]*\)')
+_MD_HEADER      = re.compile(r'^#{1,6}\s+', re.MULTILINE)
+_MULTI_SPACE    = re.compile(r'\s+')
+
+
+def _split_sentences(text: str) -> tuple[list[str], str]:
+    """Extrait les phrases complètes d'un buffer partiel.
+
+    Returns (complete_sentences, remainder).
+    """
+    parts = _SENT_RE.split(text)
+    if len(parts) <= 1:
+        return [], text
+    return parts[:-1], parts[-1]
+
+
+def _clean_for_tts(text: str) -> str:
+    """Supprime le formatage Markdown avant la synthèse vocale."""
+    text = _MD_CODE_BLOCK.sub(' ', text)
+    text = _MD_INLINE.sub(lambda m: next(g for g in m.groups() if g is not None), text)
+    text = _MD_LINK.sub(r'\1', text)
+    text = _MD_HEADER.sub('', text)
+    text = text.replace('→', ' vers ').replace('←', ' depuis ')
+    text = _MULTI_SPACE.sub(' ', text).strip()
+    return text
+
+
+# ---------------------------------------------------------------------------
+# Message processing — streaming TTS pipeline
+# ---------------------------------------------------------------------------
+
 def _process_message(user_input: str) -> None:
-    """Envoie un message au LLM et lit la réponse à voix haute."""
-    print(f"Arioch > ", end="", flush=True)
-    try:
-        reply = llm.chat(user_input)
-    except Exception as e:
-        print(f"\n[Erreur LLM] {e}")
-        return
+    """Envoie un message au LLM et lit la réponse à voix haute.
 
-    print(reply)
+    Pipeline 3 étages en parallèle :
+      [LLM stream] → sentence_queue → [TTS thread] → audio_queue → [Audio thread]
 
+    Dès qu'une phrase complète est détectée dans le stream LLM, elle part
+    immédiatement en synthèse Voxtral. L'audio est joué dès qu'il est prêt,
+    pendant que la phrase suivante est déjà en cours de synthèse.
+    """
+    SENTINEL = object()
+    sentence_q: _queue.Queue = _queue.Queue()
+    audio_q: _queue.Queue = _queue.Queue()
+
+    def tts_worker() -> None:
+        while True:
+            item = sentence_q.get()
+            if item is SENTINEL:
+                audio_q.put(SENTINEL)
+                return
+            text = _clean_for_tts(item)
+            if text:
+                try:
+                    audio_bytes = tts.text_to_speech(text)
+                    audio_q.put(audio_bytes)
+                except Exception as e:
+                    print(f"\n[Erreur TTS] {e}", flush=True)
+
+    def audio_worker() -> None:
+        while True:
+            item = audio_q.get()
+            if item is SENTINEL:
+                return
+            try:
+                audio.play_audio(item)
+            except Exception as e:
+                print(f"\n[Erreur Audio] {e}", flush=True)
+
+    tts_thread = threading.Thread(target=tts_worker, daemon=True, name="tts-worker")
+    audio_thread = threading.Thread(target=audio_worker, daemon=True, name="audio-worker")
+    tts_thread.start()
+    audio_thread.start()
+
+    print("Arioch > ", end="", flush=True)
+    buffer = ""
     try:
-        audio_bytes = tts.text_to_speech(reply)
-        audio.play_audio(audio_bytes)
+        for chunk in llm.chat_stream(user_input):
+            print(chunk, end="", flush=True)
+            buffer += chunk
+            sentences, buffer = _split_sentences(buffer)
+            for sentence in sentences:
+                sentence = sentence.strip()
+                if sentence:
+                    sentence_q.put(sentence)
     except Exception as e:
-        print(f"[Erreur TTS/Audio] {e}")
+        print(f"\n[Erreur LLM] {e}", flush=True)
+
+    # Flush any remaining text after the stream ends
+    if buffer.strip():
+        sentence_q.put(buffer.strip())
+
+    print()  # newline after full response
+    sentence_q.put(SENTINEL)
+
+    tts_thread.join()
+    audio_thread.join()
 
 
 def _handle_command(user_input: str) -> bool:
@@ -118,7 +221,7 @@ def _handle_mcp(parts: list[str]) -> None:
     print(f"\nTotal : {total} outil(s). Tapez 'mcp tools' pour les lister.\n")
 
 
-
+def _list_profiles(profiles: list) -> None:
     if not profiles:
         print("Aucun profil disponible dans profiles/")
         return