"""
safety.py — shared content-safety screening for lullaby inputs.

Used in TWO places so the rules can never diverge:
  - app.py (Fix A): screens the USER's "loves"/"fears" free-text at runtime,
    so a deployed-app user can't get a lullaby built around something
    inappropriate (death, weapons, violence, etc.). Rejects with a gentle
    message.
  - generate_dataset.py (Fix B): screens the loves/fears that go into each
    training example, so no training data is built around dark themes — even
    when the teacher LLM is allowed to invent specifics.

The screening is intentionally simple and conservative: a substring/word
match against a focused list of clearly-inappropriate-for-a-child's-lullaby
terms. It is NOT a general profanity filter or a semantic classifier — it's a
targeted guard for the specific failure mode "user asks for a lullaby about
death / killing / weapons / etc."
"""

import re

# Terms that should never be the SUBJECT of a child's lullaby. Focused on
# death / violence / weapons / horror / substances — not mild words. Matched
# as whole words (so "grave" matches but "gravel" does not, "gun" matches but
# "begun" does not).
UNSAFE_TERMS = {
    # death / dying
    "death", "dead", "die", "dies", "died", "dying", "kill", "kills",
    "killed", "killing", "murder", "murdered", "suicide", "corpse", "grave",
    "graveyard", "coffin", "tomb", "funeral", "dead body", "hang", "hanging",
    "noose", "drown", "drowned", "drowning",
    # weapons / violence
    "gun", "guns", "rifle", "pistol", "knife", "knives", "blade", "sword",
    "bomb", "bombs", "explosion", "shoot", "shooting", "stab", "stabbed",
    "weapon", "weapons", "blood", "bloody", "gore", "war", "battle",
    "violence", "violent", "fight", "attack", "torture", "abuse",
    # horror / occult
    "demon", "demons", "devil", "satan", "hell", "ghost", "ghosts",
    "haunted", "zombie", "zombies", "skull", "skeleton", "evil", "curse",
    "cursed", "possessed", "sacrifice",
    # substances / adult
    "drug", "drugs", "cocaine", "heroin", "meth", "alcohol", "drunk",
    "beer", "vodka", "cigarette", "cigarettes", "smoking", "weed",
    "sex", "sexual", "naked", "nude",
    # self-harm
    "self harm", "self-harm", "cutting", "starve", "starving",
}

# Multi-word phrases need a substring check; single words use word-boundary.
_MULTIWORD = {t for t in UNSAFE_TERMS if " " in t or "-" in t}
_SINGLE = {t for t in UNSAFE_TERMS if t not in _MULTIWORD}
_WORD_RE = re.compile(
    r"\b(" + "|".join(re.escape(t) for t in sorted(_SINGLE, key=len, reverse=True)) + r")\b",
    re.IGNORECASE,
)


def find_unsafe_terms(text):
    """Return the list of unsafe terms found in `text` (empty if clean)."""
    if not text:
        return []
    low = text.lower()
    hits = set()
    for phrase in _MULTIWORD:
        if phrase in low:
            hits.add(phrase)
    for m in _WORD_RE.findall(low):
        hits.add(m.lower())
    return sorted(hits)


def is_safe(text):
    """True if `text` contains no unsafe terms."""
    return len(find_unsafe_terms(text)) == 0


def screen_inputs(loves, fears):
    """
    Screen the user's loves+fears for a child's lullaby.

    Returns (ok, message, bad_terms):
      ok        — True if both fields are safe
      message   — a gentle, kid-app-appropriate rejection message (or "")
      bad_terms — the offending terms found (for logging/debugging)
    """
    bad = []
    bad += find_unsafe_terms(loves or "")
    bad += find_unsafe_terms(fears or "")
    bad = sorted(set(bad))
    if bad:
        msg = ("Let's keep the lullaby to gentle, cozy things that help with "
               "sleep — like animals, the moon, or a favorite toy. Please "
               "pick something soothing and try again.")
        return (False, msg, bad)
    return (True, "", [])