<?php
/**
 * Answer Steps
 *
 * @package Aisentic\Voice
 */

declare( strict_types = 1 );

namespace Aisentic\Voice;

defined( 'ABSPATH' ) || exit;

/**
 * Removes made-up how-to steps from an answer that opens a page.
 *
 * Only for a turn with no tool calls that ends with a link. It keeps:
 * - the first sentence (what page is opening),
 * - one more sentence, only if it names a setting on that page,
 * - the link.
 * All other sentences before the link are removed.
 *
 * Why: the model only knows page titles and addresses, not the fields on a
 * form. So it made up steps (for example "First give a name and a date").
 * The settings list comes from Admin_Pages 'holds' (up to five per page).
 *
 * A turn that called a tool is not changed. Its sentences may have real data.
 * It counts sentences and does not read words, so it works in Bangla too.
 *
 * @see Answer_Opening The same idea for the opening word of a refusal.
 */
class Answer_Steps {

	/**
	 * Prose, the gap, the link at the end, and any space after it.
	 *
	 * If an answer has two links, this matches the last one. The first link
	 * is treated as normal text, so it stays.
	 *
	 * @var string
	 */
	private const TRAILING_LINK = '/^(.*?)(\s*)(<a\b[^>]*>.*?<\/a>)(\s*)$/isu';

	/**
	 * Where one sentence ends and the next begins.
	 *
	 * A full stop, ?, ! or Bangla danda (।), followed by a space.
	 * The space rule keeps "9.30" and "wp-admin.php" in one piece.
	 * Speech_Chunker uses the same rule.
	 *
	 * @var string
	 */
	private const SENTENCE_END = '/(?<=[.!?\x{0964}])(?=\s)/u';

	/**
	 * Words in setting names that say nothing about the setting.
	 *
	 * Many Eventin field names start with "Enable", "Select" or "Settings".
	 * These words are ignored, so a guess like "Select a date" does not match
	 * a field called "Select Currency".
	 *
	 * @var string[]
	 */
	private const COMMON_WORDS = [ 'enable', 'disable', 'select', 'settings', 'setting', 'show', 'hide', 'allow', 'choose', 'default', 'page', 'pages', 'option', 'options', 'with', 'your', 'from', 'this', 'that' ];

	/**
	 * Drop the invented steps from an answer that opens a page.
	 *
	 * @param string $text         The whole answer, or the part of it that has
	 *                             not been spoken yet.
	 * @param int    $tool_calls   How many tools this turn asked for.
	 * @param bool   $opening_said True when the first sentence has already
	 *                             gone out to the caller as its own clip, so
	 *                             what is left here is the steps and the link.
	 * @param array  $known        Names of the settings on the page the link
	 *                             opens. A sentence naming one may stay.
	 * @return string The answer, unchanged unless a turn that looked nothing
	 *                up wrote sentences between its opening line and its link.
	 */
	public static function fix( string $text, int $tool_calls, bool $opening_said = false, array $known = [] ): string {
		// Something was looked up, so the later sentences may be real.
		if ( 0 !== $tool_calls ) {
			return $text;
		}

		$tag  = '';
		$body = $text;

		if ( preg_match( Text_To_Speech::LEADING_TONE_TAG, $text, $m ) ) {
			$tag  = $m[1];
			$body = (string) substr( $text, strlen( $m[1] ) );
		}

		// No page is opening, so there is no screen to invent steps for.
		if ( ! preg_match( self::TRAILING_LINK, $body, $parts ) ) {
			return $text;
		}

		$prose = $parts[1];
		$gap   = $parts[2];
		$link  = $parts[3] . $parts[4];

		if ( '' === trim( $prose ) ) {
			return $text;
		}

		$sentences = array_values(
			array_filter(
				array_map( 'trim', (array) preg_split( self::SENTENCE_END, $prose ) ),
				static fn( $piece ) => '' !== $piece
			)
		);

		// One sentence in front of the link is already the shape we want.
		$keep = $opening_said ? [] : array_slice( $sentences, 0, 1 );
		$next = $sentences[ count( $keep ) ] ?? '';

		// Plus one how-to sentence, when it names a setting on that page.
		if ( '' !== $next && self::names_one_of( $next, $known ) ) {
			$keep[] = $next;
		}

		if ( count( $sentences ) <= count( $keep ) ) {
			return $text;
		}

		if ( empty( $keep ) ) {
			return $tag . $link;
		}

		return $tag . implode( ' ', $keep ) . $gap . $link;
	}

	/**
	 * Check if this sentence names one of these settings.
	 *
	 * One word is enough: "Select Currency" matches a sentence with "currency".
	 * Words shorter than 4 letters and COMMON_WORDS are ignored.
	 *
	 * @param string $sentence One sentence of the answer.
	 * @param array  $known    Setting names from the page.
	 * @return bool
	 */
	private static function names_one_of( string $sentence, array $known ): bool {
		foreach ( $known as $name ) {
			foreach ( preg_split( '/[^\p{L}\p{N}]+/u', (string) $name ) as $word ) {
				if ( mb_strlen( $word ) >= 4 && ! in_array( mb_strtolower( $word ), self::COMMON_WORDS, true ) && false !== mb_stripos( $sentence, $word ) ) {
					return true;
				}
			}
		}

		return false;
	}
}