src/hooks/useAutoTTS.ts

import { useEffect, useRef } from 'react'
import { playAudioFile, edgeTts } from '../services/voice/edgeTTS.js'
import { getInitialSettings } from '../utils/settings/settings.js'
import type { RenderableMessage } from '../types/message.js'

// Map language codes to Edge TTS Chinese voices
const CHINESE_VOICES: Record<string, string> = {
  zh: 'zh-CN-XiaoxiaoNeural',
  'zh-CN': 'zh-CN-XiaoxiaoNeural',
  'zh-TW': 'zh-TW-HsiaoChenNeural',
  'zh-HK': 'zh-HK-HiuMaanNeural',
}

function resolveEdgeTTSVoice(language: string | undefined, explicitVoice: string | undefined): string {
  if (explicitVoice) return explicitVoice
  if (!language) return 'en-US-JennyNeural'
  const base = language.split('-')[0].toLowerCase()
  if (base === 'zh') return CHINESE_VOICES[language] || CHINESE_VOICES['zh-CN']
  if (base === 'ja') return 'ja-JP-NanamiNeural'
  if (base === 'ko') return 'ko-KR-SunHiNeural'
  if (base === 'fr') return 'fr-FR-DeniseNeural'
  if (base === 'de') return 'de-DE-KatjaNeural'
  if (base === 'es') return 'es-ES-ElviraNeural'
  return 'en-US-JennyNeural'
}

export function useAutoTTS(messages: RenderableMessage[], isLoading?: boolean): void {
  const triggeredIdsRef = useRef<Set<string>>(new Set())
  const pendingPlayRef = useRef<Promise<void> | null>(null)

  // Snapshot of message IDs present when the conversation first becomes "active".
  // isLoading goes true on the first user submit and stays false on resume.
  // By waiting for isLoading=true we ensure the snapshot captures the correct
  // boundary: everything before it is history (skip), everything after is new.
  const historyIdsRef = useRef<Set<string> | null>(null)
  const wasLoadingRef = useRef<boolean>(false)

  useEffect(() => {
    const settings = getInitialSettings()
    if (!settings.voiceAutoTTS || !settings.voiceEnabled) return

    // Detect session switch (mid-session /resume): if we already took a
    // history snapshot for a previous session but the incoming messages are
    // entirely different (no overlapping UUIDs), reset and re-snapshot so the
    // resumed session's history isn't re-spoken.
    if (historyIdsRef.current && messages.length > 0) {
      const isSessionSwitch = messages.every(
        m => !historyIdsRef.current!.has(m.uuid),
      )
      if (isSessionSwitch) {
        historyIdsRef.current = null
        wasLoadingRef.current = false
      }
    }

    // Capture history snapshot on the transition from idle → active, or on
    // first render with pre-loaded messages (resume path).
    if (!historyIdsRef.current && !wasLoadingRef.current) {
      if (isLoading || messages.length > 0) {
        // Normal path (isLoading=true): user just submitted their first
        // message in a fresh session. Snapshot current messages so pre-loaded
        // tool-call results / system messages are excluded from TTS.
        //
        // Resume path (messages.length>0, isLoading=false): messages were
        // pre-loaded without a loading transition (e.g. /resume,
        // --resume-session). Treat all displayed messages as history so they
        // are not re-spoken. Only subsequent real-time replies will be spoken.
        historyIdsRef.current = new Set(messages.map(m => m.uuid))
      }
    }
    wasLoadingRef.current = isLoading ?? false

    const voice = resolveEdgeTTSVoice(
      settings.voiceLanguage || settings.language,
      settings.voiceTTSVoice,
    )

    for (const msg of messages) {
      if (triggeredIdsRef.current.has(msg.uuid)) continue
      // Skip messages that were present before the conversation became active.
      if (historyIdsRef.current?.has(msg.uuid)) continue

      triggeredIdsRef.current.add(msg.uuid)

      if (msg.type !== 'assistant') continue
      // Search all content blocks for text — tool calls (e.g. WebSearch)
      // may appear as the first block with text following
      const textBlock = Array.isArray(msg.message.content)
        ? msg.message.content.find((b: any) => b?.type === 'text')
        : null
      if (!textBlock?.text?.trim()) continue

      // Strip formatting before TTS: URLs, code blocks, markdown syntax
      const text = textBlock.text
        .replace(/https?:\/\/\S+/g, '')                    // URLs
        .replace(/```[\s\S]*?```/g, '')                     // code blocks
        .replace(/(?<=^|[^*])\*{1,2}(?!\*)(.*?)\*{1,2}(?=[^*]|$)/g, '$1')  // *italic* **bold**
        .replace(/(?<=^|[^_])\_{1,2}(?!_)(.*?)\_{1,2}(?=[^_]|$)/g, '$1')  // _italic_ __bold__
        .replace(/`([^`]+)`/g, '$1')                        // inline `code`
        .replace(/~~(.*?)~~/g, '$1')                        // ~~strikethrough~~
        .replace(/\[([^\]]+)\]\([^)]+\)/g, '$1')            // [text](url)
        .replace(/^#{1,6}\s+/gm, '')                        // # headings
        .replace(/^[>\|]\s*/gm, '')                         // > blockquotes, | tables
        .replace(/(?:^|\n)[-*+]\s+/g, ' ')                  // list markers
        .replace(/\n{3,}/g, '\n\n')                         // collapse excess newlines
        .trim()

      if (!text) continue

      const run = async () => {
        const result = await edgeTts({ text, voice })
        if (result.success && result.audioPath) {
          await playAudioFile(result.audioPath)
        }
      }

      if (pendingPlayRef.current) {
        pendingPlayRef.current = pendingPlayRef.current.then(run, run).catch(() => {})
      } else {
        pendingPlayRef.current = run().catch(() => {})
      }
    }
  }, [messages, isLoading])
}