#!/usr/bin/env python3
"""從合併後的標註產生各角色的對白索引與人格卡的素材。

用法：build_pack.py <final.jsonl> <資料目錄>

每個角色輸出到 <資料目錄>/<角色代號>/：
- index.jsonl：執行期檢索用，一行一句 {id, function, emotion, mode, text}。
- source.jsonl：撰寫人格卡用，同上再附前文與冊章。

資料目錄放在 repo 之外（預設 ~/.claude/anime-companion）；原文不入版本控制。
"""
import json
import sys
from collections import Counter, defaultdict
from pathlib import Path

CHARACTERS = {"sao-asuna": "asuna", "sao-suguha": "suguha"}
MIN_LEN, MAX_LEN = 4, 80
CAPTIVE_WORDS = ("提泰妮亞", "蒂塔妮亞", "提坦妮亞", "須鄉", "奧伯龍", "鳥籠", "世界樹")
REAL_SUGUHA = ("直葉", "小直", "和人", "哥哥")


def context(rec):
    return "".join((rec["before"], rec["after"], rec["prev_par"], rec["next_par"]))


def mode_asuna(rec):
    ctx = context(rec)
    if "明日奈" in ctx:
        return "real"
    if any(w in ctx for w in CAPTIVE_WORDS):
        return "captive"
    return {1: "ingame", 3: "captive", 4: "reunion"}.get(rec["volume"])


def mode_suguha(rec):
    ctx = context(rec)
    leafa, real = ctx.count("莉法"), sum(ctx.count(w) for w in REAL_SUGUHA)
    if leafa == 0 and real == 0:
        return None
    return "leafa" if leafa >= real else "real"


MODE_FUNCS = {"asuna": mode_asuna, "suguha": mode_suguha}


def usable(rec, speaker):
    text = rec["text"]
    return (
        rec["final_speaker"] == speaker
        and rec["label"]["confidence"] == "high"
        and MIN_LEN <= len(text) <= MAX_LEN
        and not rec["suspect_duplicate"]
        and not (rec["label"]["function"] == "other" and len(text) < 6)
    )


def build(records, speaker):
    seen, rows = set(), []
    for rec in records:
        if not usable(rec, speaker) or rec["text"] in seen:
            continue
        seen.add(rec["text"])
        rows.append(rec)
    # 沒判斷出模式的句子，沿用同一章其他句子的多數模式
    modes = [MODE_FUNCS[speaker](r) for r in rows]
    by_chapter = defaultdict(Counter)
    for r, m in zip(rows, modes):
        if m:
            by_chapter[(r["volume"], r["chapter"])][m] += 1
    out = []
    for r, m in zip(rows, modes):
        if m is None and by_chapter[(r["volume"], r["chapter"])]:
            m = by_chapter[(r["volume"], r["chapter"])].most_common(1)[0][0]
        lb = r["label"]
        out.append({"id": r["id"], "function": lb["function"], "emotion": lb["emotion"], "mode": m, "text": r["text"],
                    "_ctx": {"before": r["before"], "prev_par": r["prev_par"], "volume": r["volume"], "chapter": r["chapter"]}})
    return out


def write_pack(rows, folder):
    folder.mkdir(parents=True, exist_ok=True)
    dump = lambda items: "".join(json.dumps(i, ensure_ascii=False) + "\n" for i in items)
    (folder / "index.jsonl").write_text(dump({k: v for k, v in r.items() if k != "_ctx"} for r in rows), encoding="utf-8")
    (folder / "source.jsonl").write_text(dump({**{k: v for k, v in r.items() if k != "_ctx"}, **r["_ctx"]} for r in rows), encoding="utf-8")


def main():
    final, data = Path(sys.argv[1]), Path(sys.argv[2])
    records = [json.loads(line) for line in final.read_text(encoding="utf-8").splitlines()]
    for code, speaker in CHARACTERS.items():
        rows = build(records, speaker)
        write_pack(rows, data / code)
        print(f"{code}: {len(rows)} 句；功能 {dict(Counter(r['function'] for r in rows).most_common(5))}；模式 {dict(Counter(r['mode'] for r in rows))}")


if __name__ == "__main__":
    main()