server.py

#!/usr/bin/env python3
"""
NebulaCut – Viral Shorts Backend
Analyzes YouTube videos, finds hook segments, cuts 9:16 clips.

Requirements: pip install flask flask-cors yt-dlp
System deps:  ffmpeg (must be in PATH)
"""

import json
import math
import os
import re
import struct
import subprocess
import urllib.error
import urllib.request
import uuid
from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path
import http.client
import ssl

from flask import Flask, jsonify, request, send_from_directory, Response, stream_with_context
from flask_cors import CORS
from werkzeug.utils import secure_filename

# ── Config ────────────────────────────────────────────────────────────────────
ALLOWED_EXTENSIONS = {'mp4', 'webm', 'mkv', 'avi', 'mov', 'flv', 'wmv'}
MAX_UPLOAD_SIZE = 500 * 1024 * 1024  # 500MB
def _env_int(name: str, default: int) -> int:
    raw = os.getenv(name)
    if raw is None:
        return default
    try:
        return int(raw)
    except ValueError:
        return default


def _env_float(name: str, default: float) -> float:
    raw = os.getenv(name)
    if raw is None:
        return default
    try:
        return float(raw)
    except ValueError:
        return default


_CPU_COUNT       = max(1, os.cpu_count() or 4)
HOST             = os.getenv("CLIPPER_HOST", "0.0.0.0")
PORT             = _env_int("CLIPPER_PORT", 7860)
MAX_CLIPS        = _env_int("CLIPPER_MAX_CLIPS", 3)      # fewer clips = faster end-to-end
CLIP_DURATION    = _env_float("CLIPPER_DURATION", 22)   # shorter clips encode faster
MIN_GAP_SECONDS  = _env_float("CLIPPER_MIN_GAP", 60)     # minimum spacing between clip start times
SAMPLE_RATE      = _env_int("CLIPPER_SAMPLE_RATE", 4000) # Hz for audio energy (lower = faster decode)
ENERGY_WINDOW    = _env_int("CLIPPER_ENERGY_WINDOW", 4)  # smoothing window in seconds
ENERGY_STEP      = _env_float("CLIPPER_ENERGY_STEP", 2.0)  # seconds between RMS samples (higher = faster)
GEMINI_API_KEY   = os.getenv("GEMINI_API_KEY", "").strip()
GEMINI_MODEL     = os.getenv("GEMINI_MODEL", "gemini-1.5-flash")
GEMINI_TIMEOUT   = _env_int("GEMINI_TIMEOUT", 18)
USE_GEMINI       = os.getenv("USE_GEMINI", "0").strip().lower() not in ("0", "false", "no")
DOWNLOAD_MAX_HEIGHT = _env_int("CLIPPER_DOWNLOAD_MAX_HEIGHT", 1080)
OUTPUT_HEIGHT    = _env_int("CLIPPER_OUTPUT_HEIGHT", 1920)  # 9:16 → 1080×1920
OUTPUT_WIDTH     = int(OUTPUT_HEIGHT * 9 / 16)
ENCODE_PRESET    = os.getenv("CLIPPER_ENCODE_PRESET", "ultrafast")
ENCODE_CRF       = _env_int("CLIPPER_CRF", 32)
X264_TUNE        = os.getenv("CLIPPER_X264_TUNE", "zerolatency").strip()
MAX_WORKERS      = _env_int("CLIPPER_MAX_WORKERS", min(6, _CPU_COUNT))

BASE_DIR      = Path(__file__).parent.resolve()
CLIPS_DIR     = BASE_DIR / "clips"
DOWNLOADS_DIR = BASE_DIR / "downloads"
CLIPS_DIR.mkdir(exist_ok=True)
DOWNLOADS_DIR.mkdir(exist_ok=True)
# Font path: use Linux path for Docker/HuggingFace, fallback to Windows
import platform
if platform.system() == "Windows":
    FONT_PATH = "C\\\\:/Windows/Fonts/arialbd.ttf"
else:
    FONT_PATH = "/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf"

app = Flask(__name__)
app.config['MAX_CONTENT_LENGTH'] = MAX_UPLOAD_SIZE
CORS(app)


def allowed_file(filename):
    return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS


# ── Dependency check ──────────────────────────────────────────────────────────
def check_deps():
    missing = []
    for tool, flag in [("ffmpeg", "-version"), ("ffprobe", "-version"), ("yt-dlp", "--version")]:
        r = subprocess.run([tool, flag], capture_output=True)
        if r.returncode not in (0, 1):
            missing.append(tool)
    return missing


# ── Cobalt API Download (fallback for restricted environments) ────────────────
COBALT_INSTANCES = [
    {"url": "https://api.cobalt.tools", "version": 10},
    {"url": "https://cobalt.canine.tools", "version": 10},
    {"url": "https://co.eepy.today", "version": 10},
    {"url": "https://cobalt.api.ayo.tf", "version": 10},
]


def download_via_cobalt(youtube_url: str) -> Path:
    """
    Download video using Cobalt API as fallback when yt-dlp fails
    (e.g., on HuggingFace Spaces where YouTube is blocked).
    """
    uid = uuid.uuid4().hex[:10]
    output_path = DOWNLOADS_DIR / f"{uid}.mp4"
    
    last_error = None
    
    for instance in COBALT_INSTANCES:
        api_url = instance["url"]
        
        try:
            # Cobalt API v10 format
            payload = json.dumps({
                "url": youtube_url,
                "videoQuality": "720",
                "youtubeVideoCodec": "h264",
                "filenameStyle": "basic"
            }).encode("utf-8")
            
            headers = {
                "Content-Type": "application/json",
                "Accept": "application/json",
                "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"
            }
            
            ctx = ssl.create_default_context()
            ctx.check_hostname = False
            ctx.verify_mode = ssl.CERT_NONE
            
            req = urllib.request.Request(
                api_url,
                data=payload,
                headers=headers,
                method="POST"
            )
            
            with urllib.request.urlopen(req, timeout=30, context=ctx) as response:
                response_text = response.read().decode("utf-8")
            
            try:
                data = json.loads(response_text)
            except json.JSONDecodeError as je:
                last_error = f"{api_url}: JSON error - {response_text[:100]}"
                continue
            
            # Check status
            status = data.get("status")
            
            if status == "error":
                error_code = data.get("error", {}).get("code", "unknown")
                last_error = f"{api_url}: {error_code}"
                continue
            
            # Get download URL based on status
            download_url = None
            
            if status in ("redirect", "tunnel", "stream"):
                download_url = data.get("url")
            elif status == "picker":
                picker = data.get("picker", [])
                if picker:
                    download_url = picker[0].get("url")
            elif data.get("url"):
                download_url = data["url"]
            
            if not download_url:
                last_error = f"{api_url}: No URL in response - status={status}"
                continue
            
            # Download the video file
            dl_headers = {
                "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36",
                "Accept": "*/*",
                "Accept-Encoding": "identity"
            }
            
            dl_req = urllib.request.Request(download_url, headers=dl_headers)
            
            with urllib.request.urlopen(dl_req, timeout=300, context=ctx) as resp:
                with open(output_path, "wb") as f:
                    while True:
                        chunk = resp.read(131072)  # 128KB chunks
                        if not chunk:
                            break
                        f.write(chunk)
            
            if output_path.exists() and output_path.stat().st_size > 50000:
                return output_path
            else:
                size = output_path.stat().st_size if output_path.exists() else 0
                last_error = f"{api_url}: File too small ({size} bytes)"
                if output_path.exists():
                    output_path.unlink()
                continue
                
        except urllib.error.HTTPError as e:
            last_error = f"{api_url}: HTTP {e.code}"
            continue
        except urllib.error.URLError as e:
            last_error = f"{api_url}: URL Error - {e.reason}"
            continue
        except Exception as e:
            last_error = f"{api_url}: {type(e).__name__}: {str(e)[:100]}"
            continue
    
    raise RuntimeError(f"All Cobalt instances failed. Last error: {last_error}")


# ── Step 1: Download ──────────────────────────────────────────────────────────
def download_video(youtube_url: str):
    uid = uuid.uuid4().hex[:10]
    template = str(DOWNLOADS_DIR / f"{uid}.%(ext)s")

    # Try yt-dlp first
    result = subprocess.run(
        [
            "yt-dlp",
            "--force-ipv4",
            "--geo-bypass",
            "--no-check-certificates",
            "-f", f"bestvideo[ext=mp4][height<={DOWNLOAD_MAX_HEIGHT}]+bestaudio[ext=m4a]/best[ext=mp4][height<={DOWNLOAD_MAX_HEIGHT}]/best[height<={DOWNLOAD_MAX_HEIGHT}]/best",
            "--merge-output-format", "mp4",
            "--no-playlist",
            "--no-part",
            "--concurrent-fragments", "8",
            "--write-auto-subs",
            "--write-subs",
            "--sub-langs", "en.*,en",
            "--sub-format", "vtt/srv3/json3/best",
            "-o", template,
            youtube_url,
        ],
        capture_output=True,
    )

    if result.returncode != 0:
        stderr = result.stderr.decode(errors="replace")
        # Check if it's a network error (common on HuggingFace)
        if "No address associated with hostname" in stderr or "Unable to download" in stderr:
            # Try Cobalt API as fallback
            try:
                video_path = download_via_cobalt(youtube_url)
                return video_path, [], [video_path]
            except Exception as cobalt_err:
                raise RuntimeError(f"yt-dlp failed (network blocked), Cobalt fallback also failed: {cobalt_err}")
        raise RuntimeError(f"Download failed:\n{stderr[-800:]}")

    matches = list(DOWNLOADS_DIR.glob(f"{uid}.*"))
    if not matches:
        raise RuntimeError("yt-dlp finished but produced no output file.")
    video_path = next((m for m in matches if m.suffix.lower() in (".mp4", ".webm", ".mkv")), None)
    if not video_path:
        video_path = matches[0]
    subtitle_paths = [m for m in matches if m.suffix.lower() in (".vtt", ".json3", ".srv3")]
    return video_path, subtitle_paths, matches


def _ts_to_seconds(ts: str) -> float:
    parts = ts.split(":")
    if len(parts) == 3:
        h, m, s = parts
    else:
        h, m, s = "0", parts[0], parts[1]
    return int(h) * 3600 + int(m) * 60 + float(s.replace(",", "."))


def parse_vtt(vtt_path: Path):
    cues = []
    lines = vtt_path.read_text(encoding="utf-8", errors="replace").splitlines()
    i = 0
    while i < len(lines):
        line = lines[i].strip()
        if "-->" not in line:
            i += 1
            continue
        parts = [p.strip() for p in line.split("-->")]
        if len(parts) != 2:
            i += 1
            continue
        try:
            start = _ts_to_seconds(parts[0].split(" ")[0])
            end = _ts_to_seconds(parts[1].split(" ")[0])
        except ValueError:
            i += 1
            continue
        i += 1
        text_lines = []
        while i < len(lines) and lines[i].strip():
            cleaned = re.sub(r"<[^>]+>", "", lines[i]).strip()
            if cleaned and cleaned not in ("WEBVTT",):
                text_lines.append(cleaned)
            i += 1
        text = re.sub(r"\s+", " ", " ".join(text_lines)).strip()
        if text:
            cues.append({"start": start, "end": end, "text": text})
        i += 1
    return cues


def parse_json3(json3_path: Path):
    cues = []
    try:
        data = json.loads(json3_path.read_text(encoding="utf-8", errors="replace"))
    except (OSError, json.JSONDecodeError):
        return cues

    for ev in data.get("events", []):
        segs = ev.get("segs") or []
        text = "".join(seg.get("utf8", "") for seg in segs).strip()
        text = re.sub(r"\s+", " ", text)
        if not text:
            continue
        start_ms = ev.get("tStartMs")
        if start_ms is None:
            continue
        dur_ms = ev.get("dDurationMs", 1200)
        try:
            start = float(start_ms) / 1000.0
            end = start + max(0.2, float(dur_ms) / 1000.0)
        except (TypeError, ValueError):
            continue
        cues.append({"start": start, "end": end, "text": text})
    return cues


# ── Step 2: Video info ────────────────────────────────────────────────────────
def get_video_info(video_path: Path):
    """Return (duration, width, height) via ffprobe JSON."""
    r = subprocess.run(
        ["ffprobe", "-v", "quiet", "-print_format", "json",
         "-show_format", "-show_streams", str(video_path)],
        capture_output=True,
    )
    data = json.loads(r.stdout)
    duration = float(data["format"]["duration"])
    width = height = 0
    for s in data.get("streams", []):
        if s.get("codec_type") == "video":
            width  = int(s["width"])
            height = int(s["height"])
            break
    return duration, width, height


# ── Step 3: Audio energy analysis ────────────────────────────────────────────
def extract_audio_energy(video_path: Path, duration: float):
    """
    Pipe raw mono 8kHz PCM from FFmpeg and compute per-second RMS energy.
    Returns list of (time_sec, rms) tuples.
    """
    proc = subprocess.Popen(
        [
            "ffmpeg", "-i", str(video_path),
            "-vn", "-ar", str(SAMPLE_RATE), "-ac", "1",
            "-f", "s16le", "pipe:1",
            "-loglevel", "quiet",
        ],
        stdout=subprocess.PIPE,
        stderr=subprocess.DEVNULL,
    )
    raw, _ = proc.communicate()

    if not raw:
        # Fallback: uniform energy (clips will be evenly spaced)
        return [(t, 1.0) for t in range(int(duration))]

    n = len(raw) // 2
    samples = struct.unpack(f"<{n}h", raw)

    win   = SAMPLE_RATE * ENERGY_WINDOW  # samples per window
    step  = max(SAMPLE_RATE // 2, int(SAMPLE_RATE * max(0.5, ENERGY_STEP)))
    result = []

    for i in range(0, n - win, step):
        chunk = samples[i : i + win : 16]  # subsample → faster RMS
        if not chunk:
            continue
        rms = math.sqrt(sum(int(s) * int(s) for s in chunk) / len(chunk))
        result.append((i / SAMPLE_RATE, rms))

    return result


# ── Step 4: Find hook segments ────────────────────────────────────────────────
def find_segments(energies, duration: float, n_clips: int = MAX_CLIPS):
    """
    Greedy peak selection:
    1. Smooth the RMS curve.
    2. Repeatedly pick the highest-energy moment, then
       black out a MIN_GAP_SECONDS radius around it.
    Returns list of (start, end) in seconds.
    """
    if not energies:
        step = max(60.0, (duration - CLIP_DURATION) / max(n_clips, 1))
        return [(round(i * step + 10, 2), round(i * step + 10 + CLIP_DURATION, 2))
                for i in range(n_clips) if i * step + 10 + CLIP_DURATION <= duration]

    times = [e[0] for e in energies]
    vals  = [e[1] for e in energies]

    # Smooth
    w = min(10, max(3, len(vals) // 20))
    smoothed = []
    for i in range(len(vals)):
        lo, hi = max(0, i - w), min(len(vals), i + w + 1)
        smoothed.append(sum(vals[lo:hi]) / (hi - lo))

    # Greedy selection
    used   = [False] * len(smoothed)
    peaks  = []
    # Index spacing follows CLIPPER_ENERGY_STEP (seconds between RMS samples).
    gap_idx = max(1, int(MIN_GAP_SECONDS / max(0.25, ENERGY_STEP)))

    while len(peaks) < n_clips * 2:
        best_i = max(
            (i for i in range(len(smoothed)) if not used[i]),
            key=lambda i: smoothed[i],
            default=-1,
        )
        if best_i < 0:
            break
        peaks.append(times[best_i])
        lo = max(0, best_i - gap_idx)
        hi = min(len(used), best_i + gap_idx + 1)
        for j in range(lo, hi):
            used[j] = True

    peaks.sort()
    peaks = peaks[:n_clips]

    # If we have fewer peaks than requested, fill with evenly spaced ones
    if len(peaks) < n_clips:
        step = max(60.0, (duration - CLIP_DURATION) / max(n_clips, 1))
        t = 10.0
        while len(peaks) < n_clips and t + CLIP_DURATION <= duration:
            if all(abs(t - p) >= MIN_GAP_SECONDS for p in peaks):
                peaks.append(t)
            t += step
        peaks.sort()

    # Convert to (start, end)
    segments = []
    for pt in peaks:
        start = max(0.0, pt - CLIP_DURATION * 0.25)
        end   = start + CLIP_DURATION
        if end > duration:
            end   = duration
            start = max(0.0, end - CLIP_DURATION)
        segments.append((round(start, 2), round(end, 2)))

    return segments


def build_transcript_chunks(cues, duration: float):
    if not cues:
        return []
    chunks = []
    for cue in cues[:220]:
        chunks.append({
            "start": round(float(cue["start"]), 2),
            "end": round(float(cue["end"]), 2),
            "text": cue["text"][:160],
        })
    # Keep payload bounded for Gemini
    return chunks[:180]


def _extract_json_array(raw_text: str):
    match = re.search(r"\[[\s\S]*\]", raw_text)
    if not match:
        return None
    try:
        return json.loads(match.group(0))
    except json.JSONDecodeError:
        return None


def rank_segments_with_gemini(chunks, duration: float, n_clips: int):
    if not USE_GEMINI or not GEMINI_API_KEY or not chunks:
        return None

    prompt = (
        "You are a short-form growth editor. "
        "Given transcript chunks from a long video, return ONLY a JSON array of best moments.\n"
        "Each item must have: start, end, score, reason.\n"
        f"Rules: output up to {n_clips} moments, each about {int(CLIP_DURATION)} seconds, "
        f"duration range 25-60s, no overlap over {int(MIN_GAP_SECONDS/2)}s. "
        "Prioritize hook strength, retention potential, curiosity, emotional payoff, and shareability.\n"
        f"Video duration: {round(duration,2)} seconds.\n"
        f"Transcript chunks JSON:\n{json.dumps(chunks, ensure_ascii=True)}"
    )

    payload = {
        "contents": [{"role": "user", "parts": [{"text": prompt}]}],
        "generationConfig": {"temperature": 0.2, "responseMimeType": "application/json"},
    }
    url = (
        f"https://generativelanguage.googleapis.com/v1beta/models/"
        f"{GEMINI_MODEL}:generateContent?key={GEMINI_API_KEY}"
    )

    req = urllib.request.Request(
        url,
        data=json.dumps(payload).encode("utf-8"),
        headers={"Content-Type": "application/json"},
        method="POST",
    )
    try:
        with urllib.request.urlopen(req, timeout=GEMINI_TIMEOUT) as res:
            data = json.loads(res.read().decode("utf-8", "replace"))
    except (urllib.error.URLError, TimeoutError, json.JSONDecodeError):
        return None

    text_parts = []
    for cand in data.get("candidates", []):
        content = cand.get("content", {})
        for part in content.get("parts", []):
            if isinstance(part, dict) and "text" in part:
                text_parts.append(part["text"])
    parsed = _extract_json_array("\n".join(text_parts))
    if not isinstance(parsed, list):
        return None

    segments = []
    for item in parsed:
        if not isinstance(item, dict):
            continue
        try:
            s = float(item.get("start"))
            e = float(item.get("end"))
        except (TypeError, ValueError):
            continue
        if e <= s:
            continue
        # Normalize to target clip duration
        if (e - s) < 20:
            e = s + CLIP_DURATION
        if (e - s) > 70:
            e = s + CLIP_DURATION
        if e > duration:
            e = duration
            s = max(0.0, e - CLIP_DURATION)
        segments.append((round(s, 2), round(e, 2)))

    # De-overlap and trim
    out = []
    for s, e in sorted(segments, key=lambda x: x[0]):
        if not out or abs(s - out[-1][0]) >= (MIN_GAP_SECONDS * 0.6):
            out.append((s, e))
        if len(out) >= n_clips:
            break
    return out[:n_clips] if out else None


# ── Step 5: Build 9:16 crop filter ───────────────────────────────────────────
def build_vf(width: int, height: int) -> str:
    """Return an FFmpeg -vf string that center-crops to 9:16 output size."""
    ratio = 9 / 16
    if width / height > ratio:
        # Landscape – crop left/right
        cw = int(height * ratio)
        ch = height
        cx = (width - cw) // 2
        cy = 0
    else:
        # Portrait / square – crop top/bottom
        cw = width
        ch = int(width / ratio)
        cx = 0
        cy = (height - ch) // 2

    # Force even numbers
    cw -= cw % 2
    ch -= ch % 2
    return f"crop={cw}:{ch}:{cx}:{cy},scale={OUTPUT_WIDTH}:{OUTPUT_HEIGHT}:flags=fast_bilinear"


def build_vf_pad(width: int, height: int) -> str:
    """Scale to fit inside 9:16 output preserving aspect ratio, pad remainder black."""
    # scale down to fit, keeping aspect ratio
    # then pad symmetrically to exactly 1080x1920
    return (
        f"scale={OUTPUT_WIDTH}:{OUTPUT_HEIGHT}:force_original_aspect_ratio=decrease:flags=fast_bilinear,"
        f"pad={OUTPUT_WIDTH}:{OUTPUT_HEIGHT}:(ow-iw)/2:(oh-ih)/2:black,"
        "setsar=1"
    )


def _escape_drawtext_text(text: str) -> str:
    return (
        text.replace("\\", "\\\\")
        .replace(":", "\\:")
        .replace("'", "\\'")
        .replace(",", "\\,")
        .replace("%", "\\%")
        .replace("[", "\\[")
        .replace("]", "\\]")
    )


def _hormozi_style_text(text: str) -> str:
    tokens = re.findall(r"[A-Za-z0-9']+|[^A-Za-z0-9'\s]+", text)
    emphasis = {"you", "free", "secret", "mistake", "money", "viral", "growth", "now", "never", "always"}
    out = []
    for t in tokens:
        low = t.lower()
        if low in emphasis or len(t) >= 8:
            out.append(t.upper())
        else:
            out.append(t)
    return " ".join(out).strip()


def build_caption_events(cues, clip_start: float, clip_end: float):
    if not cues:
        return []
    events = []
    overlapping = [c for c in cues if c["end"] >= clip_start and c["start"] <= clip_end]
    for cue in overlapping:
        cue_s = max(clip_start, float(cue["start"]))
        cue_e = min(clip_end, float(cue["end"]))
        text = re.sub(r"\s+", " ", cue["text"]).strip()
        if not text or cue_e <= cue_s:
            continue
        words = text.split()
        if not words:
            continue
        group_size = 4
        groups = [" ".join(words[i:i + group_size]) for i in range(0, len(words), group_size)]
        span = max(0.35, (cue_e - cue_s) / max(len(groups), 1))
        for idx, g in enumerate(groups):
            s = cue_s + idx * span - clip_start
            e = min(cue_e - clip_start, s + max(0.35, span * 0.95))
            if e - s < 0.2:
                continue
            events.append({"start": round(max(0.0, s), 2), "end": round(max(0.0, e), 2), "text": _hormozi_style_text(g)})
    return events[:50]


def build_caption_filter(events):
    if not events:
        return ""
    fs = max(38, min(72, int(OUTPUT_HEIGHT * 0.065)))
    parts = []
    for ev in events:
        txt = _escape_drawtext_text(ev["text"])
        parts.append(
            "drawtext="
            f"fontfile={FONT_PATH}:"
            f"text='{txt}':"
            "x=(w-text_w)/2:"
            "y=h*0.78:"
            f"fontsize={fs}:"
            "fontcolor=white:"
            "borderw=6:bordercolor=black@0.95:"
            "shadowx=0:shadowy=3:shadowcolor=black@0.9:"
            f"enable='between(t\\,{ev['start']}\\,{ev['end']})'"
        )
    return ",".join(parts)


# ── Step 6: Cut clip ──────────────────────────────────────────────────────────
def cut_clip(video_path: Path, start: float, end: float,
             idx: int, width: int, height: int, mode: str = "fill", cues=None) -> Path:
    name = f"short_{idx + 1}_{uuid.uuid4().hex[:6]}.mp4"
    out  = CLIPS_DIR / name
    base_vf = build_vf_pad(width, height) if mode == "pad" else build_vf(width, height)
    caption_events = build_caption_events(cues or [], start, end)
    caption_vf = build_caption_filter(caption_events)
    vf = f"{base_vf},{caption_vf}" if caption_vf else base_vf
    dur  = round(end - start, 2)

    ff = [
        "ffmpeg",
        "-ss", str(start),
        "-i", str(video_path),
        "-t", str(dur),
        "-vf", vf,
        "-c:v", "libx264",
        "-preset", ENCODE_PRESET,
    ]
    if X264_TUNE:
        ff += ["-tune", X264_TUNE]
    ff += [
        "-crf", str(ENCODE_CRF),
        "-c:a", "aac",
        "-b:a", "96k",
        "-threads", "0",
        "-movflags", "+faststart",
        "-y",
        str(out),
    ]
    r = subprocess.run(ff, capture_output=True)

    if r.returncode != 0:
        raise RuntimeError(
            f"FFmpeg failed for clip {idx + 1}: "
            f"{r.stderr.decode(errors='replace')[-600:]}"
        )
    return out


# ── Routes ────────────────────────────────────────────────────────────────────
@app.route("/api/process", methods=["POST"])
def process():
    data        = request.get_json(force=True, silent=True) or {}
    youtube_url = (data.get("youtubeUrl") or "").strip()

    if not youtube_url:
        return jsonify({"error": "youtubeUrl is required"}), 400

    n_clips = max(1, min(int(data.get("clips", MAX_CLIPS)), MAX_CLIPS))
    mode    = data.get("mode", "fill").strip().lower()
    analysis_mode = (data.get("analysisMode") or "local").strip().lower()
    if mode not in ("fill", "pad"):
        mode = "fill"
    use_gemini = analysis_mode == "gemini" and USE_GEMINI

    def generate():
        video_path = None
        subtitle_paths = []
        extra_download_files = []
        cues = []
        try:
            # 1. Download
            yield json.dumps({"phase": "download", "message": "Downloading source video..."}) + "\n"
            video_path, subtitle_paths, extra_download_files = download_video(youtube_url)
            for sub in subtitle_paths:
                suffix = sub.suffix.lower()
                if suffix == ".vtt":
                    cues.extend(parse_vtt(sub))
                elif suffix == ".json3":
                    cues.extend(parse_json3(sub))

            if not cues:
                yield json.dumps({
                    "warning": "No subtitle transcript found; captions may be limited on this video."
                }) + "\n"

            # 2. Info
            yield json.dumps({"phase": "metadata", "message": "Reading video metadata..."}) + "\n"
            duration, width, height = get_video_info(video_path)
            if duration < 20:
                yield json.dumps({"error": "Video too short (minimum 20 s)."}) + "\n"
                return
            if width == 0 or height == 0:
                yield json.dumps({"error": "Could not read video dimensions."}) + "\n"
                return

            # 3. Audio energy + viral ranking
            yield json.dumps({"phase": "analysis", "message": "Analyzing viral moments..."}) + "\n"
            segments = None
            if use_gemini:
                chunks = build_transcript_chunks(cues, duration)
                segments = rank_segments_with_gemini(chunks, duration, n_clips=n_clips)
                if not segments:
                    yield json.dumps({
                        "warning": "Gemini analysis unavailable. Falling back to local heuristic ranking."
                    }) + "\n"
            # 4. Hook segments fallback
            if not segments:
                energies = extract_audio_energy(video_path, duration)
                segments = find_segments(energies, duration, n_clips=n_clips)
            if not segments:
                yield json.dumps({"error": "No viable segments found."}) + "\n"
                return

            # Tell the frontend how many clips to expect
            yield json.dumps({"total": len(segments)}) + "\n"
            yield json.dumps({"phase": "captioning", "message": "Generating Hormozi-style captions..."}) + "\n"

            # 5. Encode ALL clips in parallel, stream each one as it finishes
            def _cut(args):
                i, s, e = args
                return i, cut_clip(video_path, s, e, i, width, height, mode, cues=cues)

            yield json.dumps({"phase": "rendering", "message": "Rendering captioned shorts..."}) + "\n"
            with ThreadPoolExecutor(max_workers=min(len(segments), max(1, MAX_WORKERS))) as pool:
                futures = {pool.submit(_cut, (i, s, e)): i
                           for i, (s, e) in enumerate(segments)}
                for future in as_completed(futures):
                    try:
                        i, clip_path = future.result()
                        yield json.dumps({
                            "clip":  f"/clips/{clip_path.name}",
                            "index": i,
                        }) + "\n"
                    except Exception as clip_err:
                        i = futures[future]
                        yield json.dumps({"warning": f"Clip {i + 1} failed: {clip_err}"}) + "\n"

        except Exception as exc:
            yield json.dumps({"error": str(exc)}) + "\n"

        finally:
            for f in extra_download_files:
                if isinstance(f, Path) and f.exists():
                    try:
                        f.unlink()
                    except OSError:
                        pass
            if video_path and isinstance(video_path, Path) and video_path.exists():
                try:
                    video_path.unlink()
                except OSError:
                    pass

    return Response(
        stream_with_context(generate()),
        mimetype="application/x-ndjson",
        headers={"X-Accel-Buffering": "no"},   # prevent proxy buffering
    )


@app.route("/api/upload", methods=["POST"])
def upload_process():
    """Process an uploaded video file instead of a YouTube URL."""
    if 'video' not in request.files:
        return jsonify({"error": "No video file provided"}), 400
    
    file = request.files['video']
    if file.filename == '':
        return jsonify({"error": "No file selected"}), 400
    
    if not allowed_file(file.filename):
        return jsonify({"error": f"Invalid file type. Allowed: {', '.join(ALLOWED_EXTENSIONS)}"}), 400

    # Get form parameters
    n_clips = max(1, min(int(request.form.get("clips", MAX_CLIPS)), 10))
    mode = request.form.get("mode", "fill").strip().lower()
    if mode not in ("fill", "pad"):
        mode = "fill"

    # Save uploaded file
    uid = uuid.uuid4().hex[:10]
    filename = secure_filename(file.filename)
    ext = filename.rsplit('.', 1)[1].lower() if '.' in filename else 'mp4'
    video_path = DOWNLOADS_DIR / f"{uid}.{ext}"
    file.save(str(video_path))

    def generate():
        cues = []
        try:
            # 1. Read metadata
            yield json.dumps({"phase": "metadata", "message": "Reading video metadata..."}) + "\n"
            duration, width, height = get_video_info(video_path)
            if duration < 20:
                yield json.dumps({"error": "Video too short (minimum 20 seconds)."}) + "\n"
                return
            if width == 0 or height == 0:
                yield json.dumps({"error": "Could not read video dimensions."}) + "\n"
                return

            # 2. Audio energy analysis
            yield json.dumps({"phase": "analysis", "message": "Analyzing viral moments..."}) + "\n"
            energies = extract_audio_energy(video_path, duration)
            segments = find_segments(energies, duration, n_clips=n_clips)
            
            if not segments:
                yield json.dumps({"error": "No viable segments found."}) + "\n"
                return

            # 3. Tell frontend how many clips to expect
            yield json.dumps({"total": len(segments)}) + "\n"
            yield json.dumps({"phase": "captioning", "message": "Generating captions..."}) + "\n"

            # 4. Encode clips in parallel
            def _cut(args):
                i, s, e = args
                return i, cut_clip(video_path, s, e, i, width, height, mode, cues=cues)

            yield json.dumps({"phase": "rendering", "message": "Rendering shorts..."}) + "\n"
            with ThreadPoolExecutor(max_workers=min(len(segments), max(1, MAX_WORKERS))) as pool:
                futures = {pool.submit(_cut, (i, s, e)): i for i, (s, e) in enumerate(segments)}
                for future in as_completed(futures):
                    try:
                        i, clip_path = future.result()
                        yield json.dumps({
                            "clip": f"/clips/{clip_path.name}",
                            "index": i,
                        }) + "\n"
                    except Exception as clip_err:
                        i = futures[future]
                        yield json.dumps({"warning": f"Clip {i + 1} failed: {clip_err}"}) + "\n"

        except Exception as exc:
            yield json.dumps({"error": str(exc)}) + "\n"

        finally:
            # Clean up uploaded file
            if video_path.exists():
                try:
                    video_path.unlink()
                except OSError:
                    pass

    return Response(
        stream_with_context(generate()),
        mimetype="application/x-ndjson",
        headers={"X-Accel-Buffering": "no"},
    )


@app.route("/clips/<path:filename>")
def serve_clip(filename):
    return send_from_directory(str(CLIPS_DIR), filename)


@app.route("/<path:filename>")
def serve_frontend_assets(filename):
    return send_from_directory(str(BASE_DIR), filename)


@app.route("/health")
def health():
    missing = check_deps()
    return jsonify({
        "status":        "ok" if not missing else "degraded",
        "missing_tools": missing,
        "gemini_ready": bool(GEMINI_API_KEY and USE_GEMINI),
        "gemini_model": GEMINI_MODEL,
    })


@app.route("/")
def root():
    return send_from_directory(str(BASE_DIR), "index.html")


# ── Entry ─────────────────────────────────────────────────────────────────────
if __name__ == "__main__":
    print("\n[>>] NebulaCut - Clipper Backend")
    print(f"     http://localhost:{PORT}\n")

    missing = check_deps()
    if missing:
        print(f"[!]  Missing: {', '.join(missing)}")
        print("     Install them or clips won't generate.\n")
    else:
        print("[OK] ffmpeg, ffprobe, yt-dlp found\n")

    app.run(host=HOST, port=PORT, debug=False, threaded=True)