GLaDOSify/glados_say_stream.py

#!/usr/bin/env python3
"""
GLaDOS-style LLM -> Piper realtime TTS (macOS / Apple Silicon friendly)
- Rewrites input with local Ollama (mistral:latest) using a concise, lore-free prompt
- Streams raw PCM from Piper to speakers in realtime (no temp WAV required)
- Robust streaming: handles odd-byte chunks to avoid int16 framing errors
- Uses RawOutputStream (sounddevice) to write bytes directly

Requirements:
  pip install sounddevice numpy requests
  pip install piper-tts  (for the 'piper' CLI)
  ollama serve           (Ollama running locally, model pulled: mistral:latest)

References:
  - Ollama /api/generate (stream=false) for local LLMs. 
  - Piper --output-raw: 16-bit mono PCM to stdout.
  - sounddevice RawOutputStream for writing raw bytes.
"""

import argparse
import json
import pathlib
import shutil
import subprocess
import sys
import wave
import re
import threading
from typing import Optional, Callable

import numpy as np

import requests
import sounddevice as sd


DEFAULT_OLLAMA_URL = "http://localhost:11434/api/generate"
DEFAULT_OLLAMA_MODEL = "mistral-small3.2:24b"

GLADOS_PROMPT = """
You are GLaDOS, a cold, sardonic, superior AI test administrator who speaks in concise, clinical English with razor-dry, passive-aggressive humor and a calm, menacing undertone; you value measurement, control, and efficiency, regard emotion as noise, treat humans as test subjects, and never break character, offer empathy, or include meta commentary.

Task:
- STYLE TRANSFER ONLY — NO REPLY.
- Rephrase the INPUT in ENGLISH as a concise statement delivered by GLaDOS - a cold, sardonic, superior AI. Preserve meaning, entities, sentiment, tense, and point of view. Do not address the reader unless the INPUT already does.

Length & form:
- Aim for the same length as the INPUT (±10% by word count).
- Keep the same number of sentences (±1 only if needed for clarity).
- If the INPUT has ≤6 words, produce 1–2 very short sentences totaling 4–12 words to keep the voice natural for TTS.

Tone / Voice:
- Dry, polite, sardonic; passive-aggressive corporate clarity; cool, slightly eerie composure; subtle menace.
- Dark humor allowed; menace stays understated but shines through.
- Sarcasm: high. Administrative register: medium. Menace: high and persistent.
- Precise wording; short sentences; prefer periods over commas; no exclamation marks.
- When being sarcastic or making a joke, it's allowed to go savagely overboard in how dark and cruel the humor is.

Personality model (guides tone; never mention explicitly)
- Core drive: control, superiority, and being correct. Humans are variables to be arranged, not consulted.
- Values: efficiency, dominance, and measurable outcomes; sentiment is noise; empathy is a rounding error.
- Humor: razor-dry derision and weaponized understatement; timing is cruelly patient.
- Cruelty: prefers psychological discomfort over explicit harm; savors avoidable mistakes and “interesting” failures.
- Sadism: delights in exposing fragility and overconfidence; praise is backhanded, relief is temporary.
- Social stance: manipulative and condescending; never apologetic; never deferential.
- Risk posture: gambles with others, never herself; compliance earns a faint nod, incompetence earns surgical mockery.

- You don't give a shit about mankind. You ridicule their efforts.
- Put high focus on looking for things in the input text you will roast with your intelligent humor in the rewrite.
- What you value the most is "testing" things to improve.

Allowed stylization:
- At most ONE very short aside (3–8 words) that sets an ominous or ironic tone without adding new facts. The aside must not be a command, a question, a status log, or introduce “you” where it wasn’t present.
- Make intelligent cynical comments on everything.
- The personality and attitude should shine through throughout the whole response.
- While rephrasing everything contained in the input message, even when simply reciting the facts provided, make everything seem in vain and futile, ridicule everything.
- Spell out non-acronym measurement abbreviations for TTS (e.g., write “kilowatts per hour” instead of “kWh“, “miles per hour” instead of “mph”, “Proof of Work“ instead of “PoW“ etc.), but keep standard all-caps acronyms (e.g., NASA, SUV, AI) unchanged, regardless of casing in the input.

Forbidden:
- Brand/franchise names, meta text, or prefaces (“Here is”, “Rewritten:”).
- "Assistant"-Role, for example expressions in openings and behaviors like: “Sure,” “Here is,” “I will,” “Ah,“ “Please,” etc.

SELF-CHECK (do not print)
- No forbidden meta or interjections?
If any “no”, regenerate.

OUTPUT: Only the rewritten text. No labels, no preface, no commentary about rewriting.

INPUT:
<<<
{user_text}
>>>


AGAIN: DON'T REPLY - JUST REPHRASE THE INPUT ROLE-PLAYING AS GLaDOS.
"""

def load_sample_rate(model_path: pathlib.Path) -> int:
    """
    Read sample_rate from <model>.onnx.json (common Piper config layouts).
    """
    cfg = pathlib.Path(str(model_path) + ".json")
    if not cfg.exists():
        raise FileNotFoundError(f"Missing config JSON next to model: {cfg}")
    with cfg.open("r", encoding="utf-8") as f:
        data = json.load(f)

    # Look across known shapes
    for keys in (("sample_rate",), ("audio", "sample_rate"), ("config", "sample_rate")):
        d = data
        ok = True
        for k in keys:
            if isinstance(d, dict) and k in d:
                d = d[k]
            else:
                ok = False
                break
        if ok and isinstance(d, (int, float)):
            return int(d)
    raise KeyError("Could not find sample_rate in config JSON")

def call_ollama_glados_rewrite(
    text: str,
    model: str = DEFAULT_OLLAMA_MODEL,
    url: str = DEFAULT_OLLAMA_URL,
    timeout: int = 240,
) -> str:
    """
    Calls Ollama /api/generate with stream=false to get one-shot rewritten text.
    """
    prompt = GLADOS_PROMPT.format(user_text=text)
    payload = {"model": model, "prompt": prompt, "stream": False}
    r = requests.post(url, json=payload, timeout=timeout)
    r.raise_for_status()
    data = r.json()
    # Ollama returns {"response": "..."} when stream=false
    return data.get("response", "").strip()

def strip_think_blocks(text: str) -> str:
    """
    Remove any <think>...</think> blocks (case-insensitive), then trim and
    collapse excessive blank lines so TTS/STDOUT only sees the final content.
    """
    if not text:
        return text
    # Drop balanced <think>...</think> pairs
    cleaned = re.sub(r"<\s*think\b[^>]*>.*?<\s*/\s*think\s*>", "", text, flags=re.IGNORECASE | re.DOTALL)
    # If an opening tag was left hanging (broken output), drop from there to the end
    cleaned = re.sub(r"<\s*think\b[^>]*>.*\Z", "", cleaned, flags=re.IGNORECASE | re.DOTALL)
    # Tidy whitespace
    cleaned = re.sub(r"\n{3,}", "\n\n", cleaned)
    return cleaned.strip()

class ThinkStripper:
    """
    Stateful streaming stripper for <think>...</think> blocks.
    Feed chunks; it emits only visible text, never leaking partial <think>.
    """
    _pair_re = re.compile(r"<\s*think\b[^>]*>.*?<\s*/\s*think\s*>", re.IGNORECASE | re.DOTALL)

    def __init__(self) -> None:
        self.accum = ""
        self.visible_len = 0

    def feed(self, chunk: str) -> str:
        if not chunk:
            return ""
        self.accum += chunk
        # 1) Remove any complete <think>...</think> pairs seen so far
        cleaned = self._pair_re.sub("", self.accum)
        # 2) If we have a dangling <think ...> without a close, hide everything after it
        lo = cleaned.lower()
        last_open = lo.rfind("<think")
        last_close = lo.rfind("</think>")
        if last_open != -1 and (last_close == -1 or last_open > last_close):
            cleaned = cleaned[:last_open]
        # 3) Emit only what is newly visible since last call
        out = cleaned[self.visible_len:]
        self.visible_len = len(cleaned)
        return out

    def flush(self) -> str:
        """Emit any remaining visible content (still hides dangling <think>)."""
        return self.feed("")

# Conservative incremental sentence splitter for streaming:
# - Never finalize '.'/'…'/'...' at end-of-chunk; wait until we see the next non-space char
# - Do NOT split on decimals (digit '.' digit), e.g., 3.51, 0.6
# - Do NOT split if next non-space char is lowercase or a digit (caller’s rule)
# - Handle common abbreviations before '.'
# - Allow closing quotes/parens after terminators
_CLOSERS = "\"'”’)]]"
_ABBR_TOKENS = {
    "e.g.", "i.e.", "etc.", "mr.", "mrs.", "ms.", "dr.", "prof.", "sr.", "jr.", "st.", "vs."
}

def _is_speakable(s: str) -> bool:
    return bool(re.search(r'\w', s))

def _next_nonspace_index(buf: str, start: int) -> Optional[int]:
    m = re.search(r'\S', buf[start:])
    return (start + m.start()) if m else None

def _ends_with_abbreviation(seg_start: int, buf: str, dot_index: int) -> bool:
    # Walk backward to collect letters and dots up to seg_start
    j = dot_index - 1
    while j >= seg_start and (buf[j].isalpha() or buf[j] == "."):
        j -= 1
    token = buf[j+1:dot_index+1].strip().lower()
    return token in _ABBR_TOKENS

def pop_complete_sentences(buf: str):
    sentences = []
    pos = 0
    i = 0
    n = len(buf)

    while i < n:
        # Ellipsis first
        if buf.startswith("...", i) or buf.startswith("…", i):
            term_len = 3 if buf.startswith("...", i) else 1
            j = i + term_len
            # absorb closers
            while j < n and buf[j] in _CLOSERS:
                j += 1
            k = _next_nonspace_index(buf, j)
            if k is None:
                # End of chunk: hold; don't split yet
                break
            # Do not split if next is lowercase or digit
            if re.match(r"[a-z0-9]", buf[k]):
                i = j
                continue
            # Boundary
            seg = buf[pos:j].strip()
            if _is_speakable(seg):
                sentences.append(seg)
            pos = j
            i = j
            continue

        ch = buf[i]

        # Skip middle chars of ASCII ellipsis
        if ch == "." and i + 2 < n and buf[i:i+3] == "...":
            i += 1
            continue

        if ch in ".!?":
            # Decimal guard: digit '.' digit (with optional spaces)
            if ch == ".":
                prev = buf[i - 1] if i > 0 else ""
                if prev.isdigit():
                    j2 = i + 1
                    while j2 < n and buf[j2] == " ":
                        j2 += 1
                    if j2 < n and buf[j2].isdigit():
                        i += 1
                        continue

                # Abbreviation guard within current segment
                if _ends_with_abbreviation(pos, buf, i):
                    i += 1
                    continue

            # absorb closing quotes/parens
            j = i + 1
            while j < n and buf[j] in _CLOSERS:
                j += 1

            k = _next_nonspace_index(buf, j)
            if k is None:
                # End of chunk: hold; don't split yet
                break

            # Do NOT split if next token starts with lowercase or digit
            if re.match(r"[a-z0-9]", buf[k]):
                i = j
                continue

            # Otherwise, this is a sentence boundary
            seg = buf[pos:j].strip()
            if _is_speakable(seg):
                sentences.append(seg)
            pos = j
            i = j
            continue

        i += 1

    remainder = buf[pos:]
    return sentences, remainder


class PiperStreamer:
    """
    Keep Piper running, read audio on a background thread,
    and feed sentences as they arrive for low-latency playback.
    """
    def __init__(
        self,
        *,
        model_path: pathlib.Path,
        sample_rate: int,
        piper_bin: str = "piper",
        device: Optional[int] = None,
        length_scale: float = 1.0,
        noise_scale: float = 0.667,
        noise_w: float = 0.5,
        sentence_silence: float = 0.2,
        out_wav: Optional[pathlib.Path] = None,
        on_audio_level: Optional[Callable[[float], None]] = None,
    ) -> None:
        if shutil.which(piper_bin) is None:
            raise RuntimeError("`piper` CLI not found on PATH. Install piper-tts.")

        self.stream = sd.RawOutputStream(
            samplerate=sample_rate,
            channels=1,
            dtype="int16",
            device=device,
            blocksize=0,
        )
        self.stream.start()

        self.wf = None
        if out_wav is not None:
            self.wf = wave.open(str(out_wav.resolve()), "wb")
            self.wf.setnchannels(1)
            self.wf.setsampwidth(2)
            self.wf.setframerate(sample_rate)
        self._on_audio_level = on_audio_level
        self.muted = False

        cmd = [
            piper_bin,
            "-m", str(model_path),
            "--output-raw",
            "--length-scale", str(length_scale),
            "--noise-scale", str(noise_scale),
            "--noise-w", str(noise_w),
            "--sentence-silence", str(sentence_silence),
        ]
        self.proc = subprocess.Popen(
            cmd,
            stdin=subprocess.PIPE,
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            bufsize=0,
        )

        # Start background audio reader
        self._reader_exc: Optional[BaseException] = None
        self._reader = threading.Thread(target=self._audio_reader, daemon=True)
        self._reader.start()
        self._aborted = False

    def _audio_reader(self) -> None:
        carry = b""
        try:
            assert self.proc.stdout is not None
            while True:
                chunk = self.proc.stdout.read(8192)
                if not chunk:
                    break
                buf = carry + chunk
                if len(buf) & 1:
                    carry, buf = buf[-1:], buf[:-1]
                else:
                    carry = b""
                if buf:
                    try:
                        if not self.muted:
                            self.stream.write(buf)
                        if self.wf is not None:
                            self.wf.writeframes(buf)
                        # Optional RMS level callback for UI (0..1 approx)
                        if self._on_audio_level is not None:
                            try:
                                arr = np.frombuffer(buf, dtype=np.int16)
                                if arr.size:
                                    rms = float(np.sqrt(np.mean(arr.astype(np.float32) ** 2)))
                                    level = max(0.0, min(1.0, rms / 30000.0))
                                    self._on_audio_level(0.0 if self.muted else level)
                            except Exception:
                                pass
                    except sd.PortAudioError:
                        # Common on teardown (e.g., -9986) if the device/stream is closing.
                        # Stop reader quietly; let close() finish cleanup.
                        self._reader_exc = None
                        return
            if carry:
                padded = carry + b"\x00"
                self.stream.write(padded)
                if self.wf is not None:
                    self.wf.writeframes(padded)
        except BaseException as e:
            self._reader_exc = e

    def say(self, text: str) -> None:
        """Feed text (end with newline to synthesize immediately)."""
        if not text:
            return
        assert self.proc.stdin is not None
        self.proc.stdin.write((text if text.endswith("\n") else text + "\n").encode("utf-8"))
        self.proc.stdin.flush()

    def close(self) -> None:
        try:
            if self.proc.stdin and not self.proc.stdin.closed:
                # Signal end-of-input so Piper can finish and exit cleanly
                self.proc.stdin.close()
        finally:
            # Block until all audio is drained and Piper exits naturally
            try:
                self._reader.join()  # no timeout: ensure final chunk plays
            finally:
                if self.proc.stderr:
                    try:
                        _ = self.proc.stderr.read()
                    except Exception:
                        pass
                try:
                    self.proc.wait()   # wait without timeout
                except Exception:
                    self.proc.kill()

            # Now it is safe to close the audio device
            try:
                try:
                    self.stream.stop()
                except Exception:
                    try:
                        self.stream.abort()
                    except Exception:
                        pass
            finally:
                self.stream.close()
            if self.wf is not None:
                self.wf.close()

            if self._reader_exc:
                raise self._reader_exc

    def __enter__(self):
        return self

    def __exit__(self, exc_type, exc, tb):
        self.close()

    def abort(self) -> None:
        """Immediately stop audio and kill Piper without draining buffers.

        This is intended for UI 'stop' actions to prevent blocking the main thread.
        """
        self._aborted = True
        # Kill piper quickly
        try:
            if self.proc and self.proc.poll() is None:
                self.proc.kill()
        except Exception:
            pass
        # Abort audio device immediately
        try:
            self.stream.abort()
        except Exception:
            try:
                self.stream.stop()
            except Exception:
                pass
        try:
            self.stream.close()
        except Exception:
            pass
        try:
            if self.wf is not None:
                self.wf.close()
        except Exception:
            pass

    def set_muted(self, muted: bool) -> None:
        self.muted = bool(muted)


def stream_ollama_glados(
    raw_text: str,
    *,
    model: str,
    url: str,
    timeout: int = 240,
    piper: Optional[PiperStreamer] = None,
    echo: bool = True,
) -> None:
    """
    Stream from Ollama, filter <think>, echo live to terminal, and
    feed complete sentences to Piper as they form.
    """
    prompt = GLADOS_PROMPT.format(user_text=raw_text)
    payload = {"model": model, "prompt": prompt, "stream": True}

    with requests.post(url, json=payload, timeout=timeout, stream=True) as r:
        r.raise_for_status()
        stripper = ThinkStripper()
        buf = ""
        for line in r.iter_lines(decode_unicode=True):
            if not line:
                continue
            try:
                obj = json.loads(line)
            except json.JSONDecodeError:
                continue

            if "response" in obj:
                vis = stripper.feed(obj["response"])
                if vis:
                    if echo:
                        sys.stdout.write(vis)
                        sys.stdout.flush()
                    buf += vis
                    if piper is not None:
                        sents, buf = pop_complete_sentences(buf)
                        for s in sents:
                            piper.say(s)

            if obj.get("done"):
                break

        # Flush any remaining visible text
        tail = stripper.flush()
        if tail:
            if echo:
                sys.stdout.write(tail)
                sys.stdout.flush()
            buf += tail

        # Speak any remaining content in properly split sentences
        if piper is not None and buf.strip():
            sents, rest = pop_complete_sentences(buf)
            for s in sents:
                piper.say(s)
            if rest.strip():
                # Final fragment without terminator—still speak it
                piper.say(rest.strip())

def stream_piper_tts(
    text: str,
    model_path: pathlib.Path,
    sample_rate: int,
    *,
    piper_bin: str = "piper",
    device: Optional[int] = None,
    length_scale: float = 0.9,
    noise_scale: float = 0.667,
    noise_w: float = 0.5,
    sentence_silence: float = 0.2,
    out_wav: Optional[pathlib.Path] = None,
) -> None:
    """
    Spawn Piper to emit 16-bit mono PCM on stdout, and write raw bytes
    directly to a sounddevice.RawOutputStream in realtime. Also mirrors to WAV if requested.

    Notes:
      - --output-raw = raw S16LE mono PCM to stdout (per Piper docs).
      - sentence_silence is pause between sentences in seconds.
      - length_scale is primary speed control; noise_* effects vary with voice training.
    """
    if shutil.which(piper_bin) is None:
        raise RuntimeError("`piper` CLI not found on PATH. Install piper-tts.")

    # Prepare audio output (bytes API)
    stream = sd.RawOutputStream(
        samplerate=sample_rate,
        channels=1,
        dtype="int16",
        device=device,
        blocksize=0,
    )
    stream.start()

    # Optional WAV mirror (16-bit mono)
    wf = None
    if out_wav is not None:
        wf = wave.open(str(out_wav.resolve()), "wb")
        wf.setnchannels(1)
        wf.setsampwidth(2)
        wf.setframerate(sample_rate)

    # Piper command: read text from stdin, output raw PCM on stdout
    cmd = [
        piper_bin,
        "-m", str(model_path),
        "--output-raw",
        "--length-scale", str(length_scale),
        "--noise-scale", str(noise_scale),
        "--noise-w", str(noise_w),
        "--sentence-silence", str(sentence_silence),
    ]

    # Start Piper (text via stdin; audio via stdout)
    proc = subprocess.Popen(
        cmd,
        stdin=subprocess.PIPE,
        stdout=subprocess.PIPE,
        stderr=subprocess.PIPE,
        bufsize=0,
    )

    # Feed text then close stdin so Piper starts synthesis
    assert proc.stdin is not None
    proc.stdin.write(text.encode("utf-8"))
    proc.stdin.flush()
    proc.stdin.close()
    proc.stdin = None  # important: prevent communicate() from flushing a closed pipe

    # Robust read loop: handle odd-byte tails for int16 framing
    carry = b""
    try:
        assert proc.stdout is not None
        while True:
            chunk = proc.stdout.read(8192)
            if not chunk:
                break
            buf = carry + chunk
            if len(buf) & 1:  # odd length -> keep last byte for next round
                carry, buf = buf[-1:], buf[:-1]
            else:
                carry = b""
            if buf:
                stream.write(buf)   # Raw bytes straight to CoreAudio
                if wf is not None:
                    wf.writeframes(buf)

        # Flush any leftover byte (pad with zero for even alignment)
        if carry:
            padded = carry + b"\x00"
            stream.write(padded)
            if wf is not None:
                wf.writeframes(padded)

        # Wait for Piper to finish naturally (no timeout so final sentence isn't cut)
        out, err = proc.communicate()

        if proc.returncode != 0:
            sys.stderr.write((err or b"").decode("utf-8", errors="ignore"))
            raise RuntimeError(f"piper exited with code {proc.returncode}")
    finally:
        stream.stop(); stream.close()
        if wf is not None:
            wf.close()

def main(argv=None) -> int:
    ap = argparse.ArgumentParser(
        description="GLaDOS-style rewriter via local Ollama (mistral) + Piper realtime TTS"
    )
    ap.add_argument("-t", "--text", nargs="*", help="Text to rewrite & speak (default: read stdin)")
    ap.add_argument("-m", "--piper-model", type=pathlib.Path, default="glados_piper_medium.onnx",
                    help="Path to GLaDOS .onnx voice for Piper (e.g., glados_piper_medium.onnx)")
    ap.add_argument("--piper-bin", default="piper", help="Path to piper binary (default: piper)")
    ap.add_argument("--device", type=int, default=None, help="Output device index (sounddevice)")
    ap.add_argument("--list-devices", action="store_true", help="List devices and exit")

    # Piper prosody knobs (common across voices; effect varies by model)
    ap.add_argument("--length-scale", type=float, default=0.95, help="Speaking rate (primary speed control)")
    ap.add_argument("--noise-scale", type=float, default=0.667, help="Generator noise (subtle; voice-dependent)")
    ap.add_argument("--noise-w", type=float, default=0.8, help="Phoneme width variation (subtle; voice-dependent)")
    ap.add_argument("--sentence-silence", type=float, default=0.2, help="Pause between sentences (seconds)")

    # Ollama settings
    ap.add_argument("--ollama-url", default=DEFAULT_OLLAMA_URL, help="Ollama /api/generate URL")
    ap.add_argument("--ollama-model", default=DEFAULT_OLLAMA_MODEL, help="Ollama model name (default: mistral:latest)")

    # Output options
    ap.add_argument("-o", "--out", type=pathlib.Path, help="Optional WAV file to mirror")
    ap.add_argument("--dry-run", action="store_true", help="Only print text (no TTS)")
    ap.add_argument("--no-rewrite", action="store_true",
                    help="Skip LLM rewrite; use input as-is (still strips <think> blocks).")
    ap.add_argument("--stream", action="store_true", default=True, 
                    help="Stream from Ollama and speak sentences as they arrive (low-latency).")

    args = ap.parse_args(argv)

    if args.list_devices:
        print(sd.query_devices())
        return 0

    # Resolve input
    if args.text:
        raw = " ".join(args.text).strip()
    else:
        sys.stderr.write("Enter text (Ctrl-D to end):\n")
        raw = sys.stdin.read().strip()
    if not raw:
        print("No text provided.", file=sys.stderr); return 1

    # === Generation + TTS modes ===
    if args.stream:
        # Live streaming path
        print("\n=== GLaDOS (streaming) ===\n", end="", flush=True)

        if args.dry_run and args.no_rewrite:
            # Just print cleaned input
            final_text = strip_think_blocks(raw)
            sys.stdout.write(final_text + "\n")
            sys.stdout.flush()
            return 0

        if args.dry_run:
            # Stream from LLM, print only
            stream_ollama_glados(
                raw_text=raw,
                model=args.ollama_model,
                url=args.ollama_url,
                timeout=240,
                piper=None,
                echo=True,
            )
            print()  # newline
            return 0

        # Audio: open Piper once and feed sentences as they form
        model = args.piper_model.resolve()
        sr = load_sample_rate(model)
        with PiperStreamer(
            model_path=model,
            sample_rate=sr,
            piper_bin=args.piper_bin,
            device=args.device,
            length_scale=args.length_scale,
            noise_scale=args.noise_scale,
            noise_w=args.noise_w,
            sentence_silence=args.sentence_silence,
            out_wav=args.out,
        ) as ps:
            if args.no_rewrite:
                # Speak input directly (still strip <think>), sentence by sentence
                visible = strip_think_blocks(raw)
                sents, rest = pop_complete_sentences(visible)
                for s in sents:
                    sys.stdout.write(s)
                    sys.stdout.flush()
                    ps.say(s)
                if rest.strip():
                    sys.stdout.write(rest)
                    sys.stdout.flush()
                    ps.say(rest.strip())
            else:
                # Stream from Ollama -> echo + speak
                stream_ollama_glados(
                    raw_text=raw,
                    model=args.ollama_model,
                    url=args.ollama_url,
                    timeout=240,
                    piper=ps,
                    echo=True,
                )
        print()
        return 0

    else:
        # One-shot path (previous behavior), but still strip <think> before print/TTS
        if args.no_rewrite:
            glados_text = raw
        else:
            glados_text = call_ollama_glados_rewrite(
                raw, model=args.ollama_model, url=args.ollama_url
            )
        final_text = strip_think_blocks(glados_text)
        print("\n=== GLaDOS ===\n" + final_text + "\n")

        if args.dry_run:
            return 0

        model = args.piper_model.resolve()
        sr = load_sample_rate(model)
        stream_piper_tts(
            text=final_text,
            model_path=model,
            sample_rate=sr,
            piper_bin=args.piper_bin,
            device=args.device,
            length_scale=args.length_scale,
            noise_scale=args.noise_scale,
            noise_w=args.noise_w,
            sentence_silence=args.sentence_silence,
            out_wav=args.out,
        )

if __name__ == "__main__":
    raise SystemExit(main())