from __future__ import annotations

import re

EOS_STOP = "<｜end▁of▁sentence｜>"
_NONTEXT_LINE_RE = re.compile(r"(?m)^[ \t]*(?:\[\s*Non(?:[ -]?Text)\s*\]|NonText)[ \t]*(?:\n|$)")


def _replace_complete_tags(text: str, opening: str, closing: str, *, preserve_content: bool) -> str:
    """Replace only complete exact tag pairs; malformed input is preserved."""
    chunks: list[str] = []
    cursor = 0
    while True:
        start = text.find(opening, cursor)
        if start < 0:
            chunks.append(text[cursor:])
            break
        end = text.find(closing, start + len(opening))
        if end < 0:
            chunks.append(text[cursor:])
            break
        chunks.append(text[cursor:start])
        if preserve_content:
            chunks.append(text[start + len(opening) : end])
        cursor = end + len(closing)
    return "".join(chunks)


def clean_model_output(text: str) -> str:
    """Remove exact model sentinels while preserving recognized content verbatim."""
    if text.endswith(EOS_STOP):
        text = text[: -len(EOS_STOP)]
    text = _replace_complete_tags(text, "<|ref|>", "<|/ref|>", preserve_content=True)
    text = _replace_complete_tags(text, "<|det|>", "<|/det|>", preserve_content=False)
    text = _NONTEXT_LINE_RE.sub("", text)
    return text.strip()


def repetition_warning(text: str) -> str | None:
    """Return a warning for obvious terminal repetition loops."""
    lines = [line.strip() for line in text.splitlines() if line.strip()]
    for index in range(2, len(lines)):
        if lines[index] == lines[index - 1] == lines[index - 2] and len(lines[index]) >= 8:
            return f"repeated line detected three times: {lines[index][:80]!r}"

    compact = re.sub(r"\s+", " ", text).strip()
    for width in (64, 48, 32):
        if len(compact) >= width * 3:
            tail = compact[-width:]
            if compact[-width * 3 :] == tail * 3:
                return f"repeated {width}-character suffix detected"
    return None