<?php

namespace AITranslate;

if (!defined('ABSPATH')) {
    exit;
}

/**
 * Remove reasoning blocks (<think>…</think>, <thinking>…</thinking>) some models put in the
 * message text. Without this the thinking would end up in the translated page.
 */
function strip_thinking_blocks(string $text): string
{
    if (stripos($text, '</think') === false) {
        return $text;
    }
    $clean = preg_replace('#<(think|thinking)\b[^>]*>.*?</\1>#is', '', $text);
    // Some models omit the opening tag and only close it.
    $clean = preg_replace('#^.*?</(think|thinking)>#is', '', (string) $clean, 1);
    return ltrim((string) $clean);
}

/**
 * Whether an HTTP error body says the model cannot run with thinking disabled.
 */
function is_mandatory_reasoning_error(string $body): bool
{
    return (bool) preg_match('/reasoning\s+is\s+mandatory|cannot\s+be\s+disabled/i', $body);
}

/**
 * Rewrite a request body for a model that cannot switch thinking off: lowest effort instead.
 *
 * @param array $data Request body that was rejected.
 * @return array{0:array,1:array}|null New body and the parameters to remember, or null when nothing can change.
 */
function mandatory_reasoning_retry(array $data): ?array
{
    if (isset($data['reasoning']['effort']) || isset($data['reasoning']['enabled'])) {
        $data['reasoning'] = ['effort' => 'low'];
        return [$data, ['reasoning' => ['effort' => 'low']]];
    }
    if (isset($data['reasoning_effort'])) {
        $data['reasoning_effort'] = 'low';
        return [$data, ['reasoning_effort' => 'low']];
    }
    return null;
}

/**
 * Estimate reasoning tokens for providers that do not report them (Anthropic's OpenAI-compatible
 * endpoint, models that write their thinking into the text): output tokens that the visible
 * answer does not account for. The answer is counted at 3 characters per token, which errs
 * towards "no thinking"; small differences are ignored.
 *
 * @param array  $data        Decoded response.
 * @param string $visibleText The answer without thinking blocks.
 */
function estimate_hidden_reasoning_tokens(array $data, string $visibleText): int
{
    $usage = $data['usage'] ?? null;
    if (!is_array($usage)) {
        return 0;
    }
    $out = $usage['completion_tokens'] ?? ($usage['output_tokens'] ?? null);
    if (!is_numeric($out)) {
        return 0;
    }
    $visible = (int) ceil(mb_strlen($visibleText) / 3);
    $hidden = (int) $out - $visible;
    // Counting characters per token is rough (JSON punctuation and accents cost more tokens). Only call it
    // thinking when the output is at least twice what the answer explains: Claude's plain answers
    // measured 40 tokens over the estimate, real thinking is a multiple of the answer.
    return ($hidden >= 60 && $hidden >= $visible) ? $hidden : 0;
}

/**
 * Whether an API error body rejects the temperature value sent with the request.
 *
 * Wording differs per provider (Anthropic says the value "may only be set to 1" or is
 * "deprecated" for newer Claude models), so any error that names temperature counts.
 * Retrying without temperature is harmless when the error was about something else.
 *
 * @param string $body Raw HTTP response body.
 * @return bool
 */
function is_temperature_error(string $body): bool
{
    $data = json_decode($body, true);
    $message = '';
    $param = '';
    if (is_array($data) && isset($data['error']) && is_array($data['error'])) {
        $message = strtolower((string) ($data['error']['message'] ?? ''));
        $param = strtolower((string) ($data['error']['param'] ?? ''));
    } else {
        $message = strtolower($body);
    }
    return $param === 'temperature' || strpos($message, 'temperature') !== false;
}