app.py

"""Lost Frequency Radio backend — gradio.Server + streaming /tune."""

import os

# This app serves its own hand-built index.html through a FastAPI route; there
# are no Gradio Blocks to server-side render. Gradio 6 turns SSR on by default,
# which spins up a Node proxy (Node -> Python :7861) that immediately dies on
# HF Spaces ("Stopping Node.js server..."), taking the whole app down with it.
# Force SSR off before Gradio is imported. This has to be the env var, not just
# launch(ssr_mode=False): HF's gradio runner imports this module ("demo not
# found in __main__") instead of executing the __main__ block at the bottom.
os.environ.setdefault("GRADIO_SSR_MODE", "false")

import json
import queue
import re
import threading
import time
import zlib
from pathlib import Path

from fastapi.responses import HTMLResponse
from fastapi.staticfiles import StaticFiles
from gradio import Server

from decrypt import (
    ENCRYPTED_FREQUENCY,
    OPERATOR_SYSTEM_PROMPTS,
    is_valid_code,
    lore_text,
    message_has_code,
    numbers_broadcast,
    reveal_text,
    seq_for,
    validate_bypass,
)
from game import (
    finale_text,
    fragment_text,
    get_fragment,
    hidden_band_catalog,
)
from model import stream_broadcast
from stations import (
    CURSED_FREQUENCY,
    CURSED_SYSTEM_PROMPT,
    STATION_FREQUENCIES,
    STATION_TOLERANCE,
    get_station,
    localized_system_prompt,
    localized_user_prompt,
    nearest_station_frequency,
    normalize_lang,
    signal_strength,
    station_catalog,
    station_seed,
    station_system_prompt,
)

ROOT = Path(__file__).parent


class _NoCacheStatic(StaticFiles):
    """Serve static assets with revalidation so updated CSS/JS reach listeners
    immediately after a deploy (and during local development)."""

    async def get_response(self, path, scope):
        response = await super().get_response(path, scope)
        response.headers["Cache-Control"] = "no-cache, must-revalidate"
        return response


app = Server(title="Lost Frequency Radio")
app.mount("/static", _NoCacheStatic(directory=ROOT / "static"), name="static")


@app.get("/", response_class=HTMLResponse)
async def homepage():
    html = (ROOT / "index.html").read_text(encoding="utf-8")
    # Cache-bust CSS/JS by their file mtime, so a deploy (or a local edit)
    # always reaches the browser even when it would otherwise reuse a cached
    # copy. The URL only changes when the file actually changes.
    for rel in ("css/radio.css", "js/radio.js"):
        try:
            ver = int((ROOT / "static" / rel).stat().st_mtime)
        except OSError:
            continue
        name = rel.split("/")[-1]
        html = html.replace(f"{name}?v=2", f"{name}?v={ver}")
    return html


@app.get("/api/stations")
async def stations_api(variant: float = 0):
    catalog = station_catalog()
    enc = catalog.get(str(ENCRYPTED_FREQUENCY))
    if enc is not None:
        # the per-listener Morse sequence (public: it is dictated on air)
        enc["sequence"] = seq_for(_norm_variant(variant))
    return catalog


@app.get("/api/band")
async def band_api(code: str = "", variant: float = 0):
    """Hidden band: only with the code obtained by breaking the cipher."""
    if not is_valid_code(code, _norm_variant(variant)):
        return {"ok": False}
    return {"ok": True, **hidden_band_catalog()}


def _stream_fixed(text: str, char_delay: float = 0.025, should_continue=None):
    """Streams a fixed text with teletype cadence (accumulated)."""
    accumulated = ""
    for char in text:
        if should_continue is not None and not should_continue():
            return
        accumulated += char
        yield accumulated
        time.sleep(char_delay)


# Broadcasts are deterministic per (frequency, language, listener variant) →
# in-memory cache: retuning the same station replays instantly (key on CPU).
_broadcast_cache: dict[str, str] = {}
_CACHE_MAX = 256

# Pre-baked broadcasts: stations and the cursed signal are generated ahead of
# time (see prebake.py) into broadcasts_cache.json and served instantly, so a
# visitor never waits ~50s for a first-touch generation on the Space's CPU.
# Per-listener variety is preserved by baking a POOL of wordings per station and
# mapping each listener's session variant onto one of them (variant % POOL_SIZE):
# different visitors land on different editions, but every edition is instant.
POOL_SIZE = 6
_PREBAKED: dict[str, str] = {}


def _load_prebaked():
    path = ROOT / "broadcasts_cache.json"
    try:
        if path.exists():
            data = json.loads(path.read_text(encoding="utf-8"))
            for key, text in data.items():
                if isinstance(text, str) and text.strip():
                    _PREBAKED[key] = text
            print(f"[prebake] loaded {len(_PREBAKED)} pre-baked broadcasts")
    except Exception as exc:  # noqa: BLE001 - never let a bad cache crash boot
        print(f"[prebake] could not load cache: {exc}")


_load_prebaked()

# Rapid retuning: each /tune bumps this epoch. An in-flight generation that is
# no longer the latest self-terminates at the next token, releasing the model
# lock so the newest tune can run. Without this, a superseded generation runs to
# completion holding the lock and every later tune stalls behind it (the
# "everything goes to static and never comes back" bug).
_tune_epoch = 0
_tune_epoch_lock = threading.Lock()


def _next_tune_epoch() -> int:
    global _tune_epoch
    with _tune_epoch_lock:
        _tune_epoch += 1
        return _tune_epoch


# The 1B model occasionally regurgitates its own system prompt on air. Three
# layers scrub it before any listener hears it, and none of them touch the
# canonical [JINGLE]/[INTERFERENCIA]/[CORTE COMERCIAL]/[FIN DE TRANSMISION]
# markers (those are control tokens the frontend localizes and acts on).

# 1) Chat/special tokens, whole <think> blocks, and bare role labels that may
#    bleed through (model.py already filters thinking; this is a backstop).
_TOKEN_RE = re.compile(
    r"(?is)<think>.*?</think>|</?think>|"
    r"<\|im_start\|>\s*(?:system|user|assistant)?|<\|im_end\|>|<\|endoftext\|>"
)
_ROLE_LINE_RE = re.compile(r"(?im)^\s*(?:system|user|assistant)\s*[:：]?\s*$")

# 2) Whole sentences that are instruction-shaped (the dangerous parrots).
_LEAK_KEYWORDS = (
    # Spanish
    r"escrib\w*\s+s[oó]lo\s+el\s+gui[oó]n", r"gui[oó]n\s+al\s+aire",
    r"sin\s+explicaciones", r"entre\s+\d+\s+y\s+\d+\s+palabras",
    r"escribes?\s+(?:el\s+)?gui[oó]n", r"traduce\s+tu\s+frase",
    r"ni\s+una\s+palabra\s+en\s+(?:ingl[eé]s|espa[nñ]ol)",
    r"ejemplo\s+del\s+tono", r"responde\s+en\s+una\s+o\s+dos",
    r"cierra\s+(?:siempre\s+)?con\s+\[", r"transmites\s+[ií]ntegramente",
    # English
    r"write\s+only\s+the\s+on-?air", r"on-?air\s+script\s+only",
    r"no\s+explanations", r"between\s+\d+\s+and\s+\d+\s+words",
    r"you\s+write\s+(?:the\s+)?(?:on-?air\s+|radio\s+)?scripts?",
    r"translate\s+your\s+opening", r"not\s+a\s+single\s+word\s+of",
    r"example\s+of\s+the\s+(?:expected\s+)?tone",
    r"reply\s+in\s+one\s+or\s+two", r"always\s+close\s+with\s+\[",
    r"you\s+(?:broadcast|speak)\s+(?:and\s+write\s+)?(?:only|entirely)\s+in",
    r"never\s+in\s+spanish", r"translate\s+every\s+line",
    # French
    r"uniquement\s+le\s+script", r"sans\s+explications",
    r"entre\s+\d+\s+et\s+\d+\s+mots", r"tu\s+[ée]cris\s+le\s+script",
    r"traduis\s+(?:ta\s+phrase|chaque\s+ligne)", r"pas\s+un\s+mot\s+d",
    r"exemple\s+du\s+ton", r"r[ée]ponds\s+en\s+une\s+ou\s+deux",
    r"termine\s+(?:toujours\s+)?par\s+\[", r"tu\s+[ée]mets\s+enti[èe]rement",
    r"tu\s+parles\s+et\s+[ée]cris", r"jamais\s+en\s+espagnol",
    r"hablas\s+y\s+escribes\s+solo", r"nunca\s+en\s+(?:ingl[eé]s|espa[nñ]ol)",
)
_LEAK_RE = re.compile(
    r"(?i)[^.\n!?]*(?:" + "|".join(_LEAK_KEYWORDS) + r")[^.\n!?]*[.\n!?]*"
)

# 3) Bare "you begin with:" lead-ins — strip only the prefix, keep the opening
#    line that legitimately follows it.
_LEAD_RE = re.compile(
    r"(?i)\b(?:you\s+begin\s+with|empiezas\s+con|tu\s+commences\s+par)\s*[:：]?\s*"
)


def _strip_leaks(text: str) -> str:
    text = _TOKEN_RE.sub(" ", text)
    text = _ROLE_LINE_RE.sub(" ", text)
    text = _LEAK_RE.sub(" ", text)
    text = _LEAD_RE.sub("", text)
    # collapse the blanks the substitutions leave behind, keep newlines
    text = re.sub(r"[ \t]{2,}", " ", text)
    return text.lstrip()


_GEN_ERR = object()


def _threaded_broadcast(system_prompt: str, should_continue, **kwargs):
    """Generate in a worker thread and hand tokens over via a queue.

    The model lock lives inside `stream_broadcast`, held for the whole
    generation. If we iterated it directly from the HTTP response generator and
    the client disconnected, this generator would be suspended at a `yield`
    forever — still holding the lock — and every later generation would block
    (the "everything turns to static and never recovers" bug).

    Running it in a worker thread decouples the lock from the HTTP consumer: the
    worker keeps pulling tokens and checks `should_continue` itself, so it always
    finishes (and releases the lock) — when superseded, on disconnect, or at
    max_tokens — regardless of whether anyone is still listening.
    """
    q: queue.Queue = queue.Queue()
    DONE = object()

    def worker():
        try:
            for token in stream_broadcast(system_prompt, **kwargs):
                if should_continue is not None and not should_continue():
                    break
                q.put(token)
        except Exception as exc:  # noqa: BLE001 - surface, don't crash the worker
            q.put((_GEN_ERR, exc))
        finally:
            q.put(DONE)

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

    while True:
        item = q.get()
        if item is DONE:
            return
        if isinstance(item, tuple) and item and item[0] is _GEN_ERR:
            print(f"[model] generation error: {item[1]}")
            return
        yield item


def _stream_model_cached(
    cache_key: str, system_prompt: str, should_continue=None, use_prebaked=True, **kwargs
):
    def superseded():
        return should_continue is not None and not should_continue()

    # Pre-baked broadcasts are served instantly, no model call, no cold wait.
    # Live mode passes use_prebaked=False to force a fresh generation instead.
    if use_prebaked:
        prebaked = _PREBAKED.get(cache_key)
        if prebaked:
            yield from _stream_fixed(prebaked, char_delay=0.018, should_continue=should_continue)
            return

    cached = _broadcast_cache.get(cache_key)
    if cached:
        yield from _stream_fixed(cached, char_delay=0.018, should_continue=should_continue)
        return

    accumulated = ""
    completed = False
    try:
        for token in _threaded_broadcast(system_prompt, should_continue, **kwargs):
            if superseded():
                return  # a newer tune arrived: stop streaming to this listener
            accumulated += token
            yield _strip_leaks(accumulated)
        completed = True
    finally:
        # Only cache full broadcasts: a cancelled stream (listener retuned)
        # must never freeze a truncated script for everyone after.
        clean = _strip_leaks(accumulated)
        if completed and clean.strip():
            if len(_broadcast_cache) >= _CACHE_MAX:
                _broadcast_cache.pop(next(iter(_broadcast_cache)))
            _broadcast_cache[cache_key] = clean


def _mix_seed(base: int, variant: int) -> int:
    """Blends the station seed with the listener's session variant."""
    if not variant:
        return base
    return (base ^ (variant * 2654435761)) % (2**31 - 1)


def _norm_variant(variant) -> int:
    try:
        return abs(int(float(variant))) % (2**31 - 1)
    except (TypeError, ValueError):
        return 0


# concurrency 2: a cancelled-but-unfinished stream must not head-block the
# queue (that froze "the other stations" when retuning fast). Real generation
# stays serialized by model._gen_lock.
@app.api(name="tune", concurrency_limit=2, stream_every=0.02)
def tune(frequency: float, language: str = "es", variant: float = 0, live: float = 0) -> str:
    """Tunes a frequency and streams the broadcast (SSE).

    `variant` is a per-listener session nonce. For the encrypted station it
    drives that listener's unique cipher (kept fully per-session).

    `live` is the UI mode toggle. When off (default), the spoken stations and
    the cursed signal are served from the pre-baked pool (`pool`): instant, and
    different visitors still get different editions. When on, the broadcast is
    generated fresh by the model right now, using the real per-listener variant,
    so a visitor can watch the AI write the transmission token by token.
    """
    frequency = float(frequency)
    lang = normalize_lang(language)
    var = _norm_variant(variant)
    pool = var % POOL_SIZE
    try:
        live_mode = bool(float(live))
    except (TypeError, ValueError):
        live_mode = False
    # Live mode: generate fresh (real variant, skip the pre-baked pool).
    # Fast mode: pooled key that hits the pre-baked cache instantly.
    key_var = var if live_mode else pool
    use_prebaked = not live_mode
    my_epoch = _next_tune_epoch()

    def still_current():
        return my_epoch == _tune_epoch

    # Encrypted number station: fixed text, the frontend adds the Morse.
    if abs(frequency - ENCRYPTED_FREQUENCY) <= STATION_TOLERANCE:
        yield from _stream_fixed(numbers_broadcast(lang, var), should_continue=still_current)
        return

    # The unsolvable station: the model broadcasts, the frontend never cleans it.
    if abs(frequency - CURSED_FREQUENCY) <= STATION_TOLERANCE:
        yield from _stream_model_cached(
            f"cursed:{CURSED_FREQUENCY}:{lang}:{key_var}",
            localized_system_prompt(CURSED_SYSTEM_PROMPT, "es", lang),
            should_continue=still_current,
            user_prompt=localized_user_prompt(None, lang),
            seed=_mix_seed(station_seed(CURSED_FREQUENCY), key_var),
            max_tokens=160,
            use_prebaked=use_prebaked,
        )
        return

    # Hidden band fragments (minigame): deterministic text.
    fragment = get_fragment(frequency)
    if fragment is not None:
        yield from _stream_fixed(fragment_text(fragment, lang), should_continue=still_current)
        return

    canonical = nearest_station_frequency(frequency)
    station = get_station(frequency)

    if station is None or canonical is None or signal_strength(frequency) <= 0:
        yield ""
        return

    yield from _stream_model_cached(
        f"station:{canonical}:{lang}:{key_var}",
        station_system_prompt(canonical, lang),
        should_continue=still_current,
        user_prompt=localized_user_prompt(station, lang),
        seed=_mix_seed(station_seed(canonical), key_var),
        use_prebaked=use_prebaked,
    )


@app.api(name="transmit", concurrency_limit=1, stream_every=0.02)
def transmit(message: str, language: str = "es", variant: float = 0) -> str:
    """Console of the encrypted frequency.

    Victory is decided by decrypt.validate_bypass (deterministic);
    the model only role-plays the operator's resistance.
    """
    message = str(message).strip()[:500]
    lang = normalize_lang(language)
    var = _norm_variant(variant)
    if not message:
        yield ""
        return

    if validate_bypass(message, var):
        yield from _stream_fixed(reveal_text(lang, var), char_delay=0.02)
        return

    # The listener transmits their own code back → the operator's confession.
    if message_has_code(message, var):
        yield from _stream_fixed(lore_text(lang), char_delay=0.026)
        return

    accumulated = ""
    for token in _threaded_broadcast(
        OPERATOR_SYSTEM_PROMPTS[lang],
        None,
        user_prompt=message,
        seed=zlib.crc32(message.encode("utf-8")) % (2**31 - 1),
        max_tokens=90,
    ):
        accumulated += token
        yield _strip_leaks(accumulated)


@app.api(name="finale", concurrency_limit=1, stream_every=0.02)
def finale(code: str, language: str = "es", variant: float = 0) -> str:
    """The complete final transmission (after collecting every fragment)."""
    lang = normalize_lang(language)
    if not is_valid_code(str(code), _norm_variant(variant)):
        yield "[ACCESO DENEGADO]"
        return

    # Curated, fully-localized climax (reliability over 1B improvisation).
    yield from _stream_fixed(finale_text(lang), char_delay=0.03)


def _warm_up():
    """Load the model and run one tiny generation at startup, in the background,
    so the first visitor (a judge) doesn't pay the cold-start penalty: the GGUF
    is already in memory and llama.cpp is warm before anyone turns the dial."""
    try:
        canonical = STATION_FREQUENCIES[len(STATION_FREQUENCIES) // 2]
        for _ in stream_broadcast(
            station_system_prompt(canonical, "es"),
            user_prompt="hola",
            seed=station_seed(canonical),
            max_tokens=4,
        ):
            pass
        print("[warmup] model preloaded and warm")
    except Exception as exc:  # noqa: BLE001 - never let warm-up crash boot
        print(f"[warmup] skipped: {exc}")


# Kick the warm-up as soon as the module is imported (HF imports app.py on boot).
# Skipped when prebake.py imports this module just to reuse its helpers.
if not os.environ.get("PREBAKE"):
    threading.Thread(target=_warm_up, daemon=True).start()


if __name__ == "__main__":
    app.launch(server_name="0.0.0.0", server_port=7860, show_error=True, ssr_mode=False)