cat << 'EOF' > /home/santos/.openclaw/workspace/scripts/jervasio-manual.py
#!/usr/bin/env python3
import time, json, subprocess, sys, re, queue, threading
import numpy as np
import sounddevice as sd
import requests
from faster_whisper import WhisperModel

# Configurações
PIPER_EXE = "/home/santos/.openclaw/workspace/jervasio_env/bin/piper"
PIPER_MODEL = "/home/santos/.openclaw/workspace/models/pt_PT-tugao-medium.onnx"
WHISPER_MODEL_SIZE = "base"
RECORD_SECONDS = 6

GATEWAY_URL = "http://127.0.0.1:18789/v1/responses"
GATEWAY_TOKEN = "55b6988ba106c77aa19159079166fe0f810fe397bc117f81"

tts_queue = queue.Queue()

print("\n🤖 Jervásio: Modelos prontos. À escuta...", flush=True)
try:
    whisper = WhisperModel(WHISPER_MODEL_SIZE, device="cpu", compute_type="int8")
except Exception as e:
    print(f"❌ Erro Whisper: {e}", flush=True); sys.exit(1)

def speak(text):
    t_start = time.time()
    clean_text = text.replace('"', "").replace("\n", " ").replace("*", "").strip()
    if not clean_text:
        return
    cmd = f'echo "{clean_text}" | {PIPER_EXE} --model {PIPER_MODEL} --output_raw | paplay --raw --rate=22050 --channels=1 --format=s16le'
    subprocess.run(cmd, shell=True)
    print(f"   ⏱️  [TTS (Piper)]: {time.time() - t_start:.2f}s", flush=True)

def tts_worker():
    while True:
        chunk = tts_queue.get()
        if chunk is None: break
        try:
            speak(chunk)
        finally:
            tts_queue.task_done()

threading.Thread(target=tts_worker, daemon=True).start()

def pop_speakable_chunk(buf: str):
    m = re.search(r"(.+?[.!?;:])(\s|$)", buf, re.S)
    if not m: return None, buf
    chunk = m.group(1).strip()
    rest = buf[m.end():].lstrip()
    return chunk, rest

def send_via_gateway_stream(text):
    headers = {
        "Authorization": f"Bearer {GATEWAY_TOKEN}",
        "Content-Type": "application/json",
        "x-openclaw-agent-id": "main",
        "x-openclaw-session-key": "agent:main:main",
        "x-openclaw-message-channel": "webchat"
    }
    payload = {"model": "openclaw", "input": text, "stream": True}

    full_text, tts_buffer = "", ""
    event_type, primeira_palavra = None, False
    t_start = time.time()

    print("⏳ A enviar para o Gateway...", flush=True)
    try:
        with requests.post(GATEWAY_URL, headers=headers, json=payload, stream=True) as r:
            r.raise_for_status()
            print("🤖 Jervásio: ", end="", flush=True)
            for line in r.iter_lines(decode_unicode=True):
                if line == "" or line.startswith("event: ") or not line.startswith("data: "):
                    continue
                data = line[6:]
                if data == "[DONE]": break
                try:
                    obj = json.loads(data)
                    delta = obj.get("delta", "")
                    if delta:
                        if not primeira_palavra:
                            print(f"\n   ⏱️  [Cérebro TTFT]: {time.time() - t_start:.2f}s", flush=True)
                            print("🤖 ", end="", flush=True)
                            primeira_palavra = True
                        full_text += delta
                        tts_buffer += delta
                        print(delta, end="", flush=True)
                        while True:
                            chunk, tts_buffer = pop_speakable_chunk(tts_buffer)
                            if not chunk: break
                            tts_queue.put(chunk)
                except: pass
    except Exception as e: print(f"\n❌ Erro: {e}")
    if tts_buffer.strip(): tts_queue.put(tts_buffer.strip())
    return full_text.strip()

while True:
    try:
        input("\n[ ENTER PARA FALAR ]")
        t_rec_start = time.time()
        print(f"🎤 Gravando ({RECORD_SECONDS}s)...", flush=True)
        audio = sd.rec(int(RECORD_SECONDS * 16000), samplerate=16000, channels=1, dtype="float32")
        sd.wait()
        print(f"   ⏱️  [Gravação]: {time.time() - t_rec_start:.2f}s", flush=True)

        t_stt_start = time.time()
        segments, _ = whisper.transcribe(np.squeeze(audio), language="pt", initial_prompt="Jervásio")
        user_text = " ".join([s.text for s in segments]).strip()
        print(f"   ⏱️  [STT (Whisper)]: {time.time() - t_stt_start:.2f}s", flush=True)

        if user_text:
            print(f"👤 Tu: {user_text}", flush=True)
            send_via_gateway_stream(user_text)
            tts_queue.join()
    except KeyboardInterrupt:
        print("\nAté breve!"); break
EOF