"""Deterministic Oracle engine over a JSON attribute database.

- load_items(category)            -> list of {name, category, attributes}
- filter_candidates(items, facts) -> items consistent with facts (yes/no/unknown)
- choose_attribute(items, asked)  -> attribute that best splits the set (info gain)

All pure Python — filtering is always factually correct (the JSON is the truth).
"""

from __future__ import annotations

import json
import os
import shutil

HERE = os.path.dirname(os.path.abspath(__file__))

# Read-only seed shipped with the app.
SEED_DIR = os.path.join(HERE, "data")
# Where the live DB lives. Point ORACLE_DATA_DIR at a mounted HF Storage Bucket
# (e.g. /data) so items taught during play persist across Space restarts.
DATA_DIR = os.environ.get("ORACLE_DATA_DIR") or SEED_DIR

# category -> json filename
_FILES = {"animal": "animals.json", "fruit": "fruits.json", "vegetable": "vegetables.json"}


def _seed_if_missing(path: str, fname: str) -> None:
    """On first run against an empty bucket, copy the bundled seed DB over."""
    if os.path.exists(path):
        return
    seed = os.path.join(SEED_DIR, fname)
    if os.path.exists(seed) and os.path.abspath(path) != os.path.abspath(seed):
        os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
        shutil.copyfile(seed, path)
        print(f"[engine] seeded {fname} -> {path}", flush=True)

# Fallback question phrasing per attribute (used if the LLM question-maker is
# unavailable or returns SKIP). Mirrors the attribute keys in the JSON.
ATTR_QUESTIONS = {
    # animal
    "mammal": "Is it a mammal?",
    "bird": "Is it a bird?",
    "water": "Does it live mainly in water?",
    "carnivore": "Does it mainly eat meat?",
    "big": "Is it bigger than a human?",
    "domestic": "Is it a pet or farm animal?",
    "can_fly": "Can it fly?",
    "stripes": "Does it have stripes?",
    "horns": "Does it have horns or antlers?",
    # fruit
    "red": "Is it red?",
    "sweet": "Is it sweet?",
    "has_pit": "Does it have a hard pit or stone inside?",
    "tree": "Does it grow on a tree?",
    "tropical": "Is it a tropical fruit?",
    "peel": "Do you usually peel it before eating?",
    # vegetable
    "root": "Is it a root vegetable?",
    "leafy": "Is it a leafy vegetable?",
    "green": "Is it green?",
    "underground": "Does it grow underground?",
    "raw": "Is it usually eaten raw?",
    "long": "Is it long in shape?",
    "spicy": "Is it spicy or pungent?",
    "starchy": "Is it starchy?",
    # animal (added)
    "climbs": "Can it climb trees well?",
    "hops": "Does it hop or jump to move around?",
    "wool": "Does it have wool?",
    "hump": "Does it have a hump on its back?",
    "black_white": "Is it black and white?",
    "long_tail": "Does it have a long tail?",
    "pack": "Does it live in a group or pack?",
    # fruit (added)
    "small": "Is it small or bite-sized?",
    "seeds_outside": "Does it have seeds on the outside?",
    "hard_shell": "Does it have a hard shell?",
    "spiky": "Does it have a spiky skin?",
    # vegetable (added)
    "round": "Is it round in shape?",
    "white": "Is it white in colour?",
    "cooked": "Is it usually cooked before eating?",
    "pod": "Does it grow in a pod?",
    "thin": "Is it thin and slender?",
    "knobbly": "Is it knobbly or bumpy?",
}


# cache keyed by file path -> (mtime, items). Reloads automatically when the
# JSON changes, so hand-edits are picked up on the next game with NO restart.
_CACHE: dict = {}


def load_items(category: str) -> tuple:
    """Load the JSON DB for a category. Cached, but auto-reloads if the file
    has changed on disk (compares modification time)."""
    fname = _FILES.get(category)
    if not fname:
        return ()
    path = os.path.join(DATA_DIR, fname)
    _seed_if_missing(path, fname)
    try:
        mtime = os.path.getmtime(path)
    except OSError:
        return ()
    cached = _CACHE.get(path)
    if cached and cached[0] == mtime:
        return cached[1]
    with open(path, "r", encoding="utf-8") as f:
        items = tuple(json.load(f))
    _CACHE[path] = (mtime, items)
    return items


def _cache_clear() -> None:
    """Drop the in-memory cache (used after writing the DB programmatically)."""
    _CACHE.clear()


load_items.cache_clear = _cache_clear  # keep the old API used by discovery/tests


def filter_candidates(items: list, facts: list) -> list:
    """Keep items consistent with the facts.

    facts: [{"attribute": str, "answer": "yes"|"no"|"unknown"}].
    Unknown answers (and missing attribute values) never eliminate an item.
    """
    out = list(items)
    for fact in facts:
        attr = fact.get("attribute")
        ans = str(fact.get("answer", "")).strip().lower()
        if not attr or ans not in ("yes", "no"):
            continue  # unknown / not-sure -> no filtering
        expected = ans == "yes"
        out = [it for it in out
               if it["attributes"].get(attr) is None or it["attributes"].get(attr) == expected]
    return out


def choose_attribute(category: str, items: list, asked: list) -> str | None:
    """Pick the unused attribute whose yes/no split is closest to 50/50."""
    asked = set(asked or [])
    best, best_score = None, -1
    # consider attributes that actually appear in the data, in a stable order
    seen = []
    for it in items:
        for k in it["attributes"]:
            if k not in seen:
                seen.append(k)
    for attr in seen:
        if attr in asked:
            continue
        yes = sum(1 for it in items if it["attributes"].get(attr) is True)
        no = sum(1 for it in items if it["attributes"].get(attr) is False)
        if yes == 0 or no == 0:
            continue
        score = min(yes, no)
        if score > best_score:
            best, best_score = attr, score
    return best