"""Diagnostic: why don't [laughs]/[crying] audio tags render?

Answers three questions with measurements, not theories:
  1. Is `eleven_v3` actually available to this account? (If not, the API may be
     serving a v2 model that ignores audio tags entirely.)
  2. Does voice_settings suppress tags? v3 documents stability as discrete
     0.0/0.5/1.0, and higher stability = less responsive to audio tags. We
     currently send 0.40-0.50 + style + use_speaker_boost.
  3. Does tag POSITION (head vs tail) or tag CHOICE ([laughs] vs
     [laughs harder]) decide whether it fires?

Method: synthesize the same sentence under N configs and measure duration.
A rendered laugh/sob ADDS audible time. Control (no tag) is the baseline;
any config whose duration ~= control did not render its tag.

Run:
    ELEVENLABS_API_KEY=sk_... <venv>/bin/python scripts/_tts_probe.py
Outputs mp3s to scripts/_probe_out/ so you can listen to confirm.
"""
import os
import subprocess
import sys
from pathlib import Path

OUT = Path(__file__).resolve().parent / "_probe_out"
VOICE = "TbMNBJ27fH2U0VgpSNko"  # joy-female, same voice the laughter clips use
SAD_VOICE = "m3yAHyFEFKtbCIM5n7GF"  # sadness-female

LAUGH_BASE = "가끔은 밥 차리는 게 하기 싫을 때가 있잖아요"
SOB_BASE = "점수가 반이나 잘렸어"

CUR = {"stability": 0.45, "similarity_boost": 0.75, "style": 0.30, "use_speaker_boost": True}
CREATIVE = {"stability": 0.0, "similarity_boost": 0.75}   # v3 "Creative": most tag-responsive
NATURAL = {"stability": 0.5, "similarity_boost": 0.75}    # v3 "Natural"

# (label, text, voice, voice_settings, model)
PROBES = [
    ("01_control_notag",      f"{LAUGH_BASE}.",                  VOICE, CUR,      "eleven_v3"),
    ("02_head_laughs_cur",    f". [laughs] {LAUGH_BASE}.",       VOICE, CUR,      "eleven_v3"),
    ("03_head_laughs_creative", f". [laughs] {LAUGH_BASE}.",     VOICE, CREATIVE, "eleven_v3"),
    ("04_head_laughsharder_cur", f". [laughs harder] {LAUGH_BASE}.", VOICE, CUR,  "eleven_v3"),
    ("05_tail_laughs_cur",    f"{LAUGH_BASE} [laughs].",         VOICE, CUR,      "eleven_v3"),
    ("06_tail_laughs_creative", f"{LAUGH_BASE} [laughs].",       VOICE, CREATIVE, "eleven_v3"),
    ("07_head_laughs_natural", f". [laughs] {LAUGH_BASE}.",      VOICE, NATURAL,  "eleven_v3"),
    ("08_head_laughs_v2",     f". [laughs] {LAUGH_BASE}.",       VOICE, CUR,      "eleven_multilingual_v2"),
    ("09_sob_control",        f"{SOB_BASE}.",                    SAD_VOICE, CUR,      "eleven_v3"),
    ("10_sob_tail_cur",       f"{SOB_BASE} [crying].",           SAD_VOICE, CUR,      "eleven_v3"),
    ("11_sob_tail_creative",  f"{SOB_BASE} [crying].",           SAD_VOICE, CREATIVE, "eleven_v3"),
    ("12_sob_double_tag",     f". [sorrowful] {SOB_BASE} [crying].", SAD_VOICE, CUR,  "eleven_v3"),
]


def dur(p: Path) -> float:
    try:
        r = subprocess.run(
            ["ffprobe", "-v", "error", "-show_entries", "format=duration",
             "-of", "csv=p=0", str(p)],
            capture_output=True, text=True, check=True)
        return float(r.stdout.strip())
    except Exception:
        return float("nan")


def main():
    key = os.getenv("ELEVENLABS_API_KEY")
    if not key:
        sys.exit("ELEVENLABS_API_KEY not set")
    from elevenlabs.client import ElevenLabs
    client = ElevenLabs(api_key=key)

    print("=" * 72)
    print("Q1: which models does this account actually have?")
    print("=" * 72)
    try:
        for m in client.models.list():
            mid = getattr(m, "model_id", "?")
            if "v3" in mid.lower() or "multilingual" in mid.lower() or "turbo" in mid.lower():
                print(f"  {mid:34s} tts={getattr(m,'can_do_text_to_speech','?')} "
                      f"name={getattr(m,'name','?')}")
    except Exception as e:
        print(f"  [models.list failed] {e}")

    OUT.mkdir(exist_ok=True)
    print()
    print("=" * 72)
    print("Q2/Q3: does the tag add audible time? (control = no tag)")
    print("=" * 72)
    results = {}
    for label, text, voice, vs, model in PROBES:
        p = OUT / f"{label}.mp3"
        try:
            audio = client.text_to_speech.convert(
                voice_id=voice, text=text, model_id=model,
                output_format="mp3_44100_128", voice_settings=vs)
            with p.open("wb") as f:
                for c in audio:
                    f.write(c)
            d = dur(p)
            results[label] = d
            print(f"  {label:28s} {d:6.2f}s   model={model:24s} stab={vs.get('stability')}")
        except Exception as e:
            print(f"  {label:28s} FAILED: {str(e)[:90]}")

    print()
    print("=" * 72)
    print("VERDICT (delta vs control; a rendered laugh/sob should add >0.4s)")
    print("=" * 72)
    for ctrl, group in (("01_control_notag", [k for k in results if k.startswith(("02", "03", "04", "05", "06", "07", "08"))]),
                        ("09_sob_control", [k for k in results if k.startswith(("10", "11", "12"))])):
        if ctrl not in results:
            continue
        base = results[ctrl]
        print(f"\n  control {ctrl} = {base:.2f}s")
        for k in sorted(group):
            d = results[k] - base
            verdict = "TAG FIRED" if d > 0.4 else "no effect"
            print(f"    {k:28s} {results[k]:6.2f}s  {d:+6.2f}s  {verdict}")
    print(f"\nListen to the mp3s in {OUT} to confirm.")


if __name__ == "__main__":
    main()