diff --git a/addon/appModules/_chatCache.py b/addon/appModules/_chatCache.py
index 4bfee75..5328f9b 100644
--- a/addon/appModules/_chatCache.py
+++ b/addon/appModules/_chatCache.py
@@ -26,6 +26,10 @@
 _tempPath = None
 _chatRoomName = None
 _lastMatchedIdx = None
+# Reply context for the most recent successful lookup, or None when the
+# matched bubble was not a reply. Populated by lookupMessage(); consumed
+# by the left-arrow handler so users can hear the quoted original.
+_lastReplyInfo = None
 
 # _messageIndexMap[i] = 1-based message number for position i (0 for date rows)
 _messageIndexMap = []
@@ -53,13 +57,14 @@ def setCache(messages, tempPath, chatRoomName):
 	cache always reflects a single chat room.  Builds _messageIndexMap and
 	_messageDateGroups using the same counting rules as MessageReaderDialog.
 	"""
-	global _messages, _tempPath, _chatRoomName, _lastMatchedIdx
+	global _messages, _tempPath, _chatRoomName, _lastMatchedIdx, _lastReplyInfo
 	global _messageIndexMap, _messageDateGroups
 	clearCache()
 	_messages = list(messages or [])
 	_tempPath = tempPath
 	_chatRoomName = chatRoomName
 	_lastMatchedIdx = None
+	_lastReplyInfo = None
 
 	msgIdx = 0
 	lastDateIdx = -1
@@ -81,13 +86,14 @@ def setCache(messages, tempPath, chatRoomName):
 
 def clearCache():
 	"""Reset cache state and remove the temp export file if present."""
-	global _messages, _tempPath, _chatRoomName, _lastMatchedIdx
+	global _messages, _tempPath, _chatRoomName, _lastMatchedIdx, _lastReplyInfo
 	global _messageIndexMap, _messageDateGroups
 	path = _tempPath
 	_messages = []
 	_tempPath = None
 	_chatRoomName = None
 	_lastMatchedIdx = None
+	_lastReplyInfo = None
 	_messageIndexMap = []
 	_messageDateGroups = []
 	if path:
@@ -252,6 +258,115 @@ def _formatMessage(msg):
 	return f"{name} {content} {timeStr}".strip()
 
 
+def _detectReplyNames(ocrText):
+	"""Identify both names when the OCR shows a reply preview.
+
+	Reply bubbles show the actual sender's name first (above the bubble),
+	then a quote block containing the quoted user's name and message
+	preview, then the actual reply content. OCR captures both names, and
+	the longer quoted preview otherwise wins on content overlap alone.
+
+	Returns (replySender, quotedSender) when 2+ distinct cached sender
+	names appear in the OCR text (reply pattern). Names are ordered by
+	their position in the OCR text — the earliest is the actual sender,
+	the next is the quoted user being replied to. Returns (None, None)
+	when the OCR doesn't look like a reply.
+	"""
+	if not ocrText:
+		return None, None
+	occurrences = []
+	seen = set()
+	for msg in _messages:
+		if msg.get("type") != "message":
+			continue
+		name = msg.get("name", "")
+		if not name or name in seen:
+			continue
+		# Require the name at the start of a line, optionally preceded by
+		# non-letter glyphs (e.g. the "0 " quote-indicator icon LINE
+		# renders before quoted-user names in reply bubbles), and not
+		# immediately followed by another CJK/Latin letter.  This:
+		#   - lets quoted names like "0 王昱涵" still match (real LINE OCR);
+		#   - blocks names mentioned mid-sentence (e.g. "感謝Bob你的幫助"
+		#     — 感謝 is CJK so the non-letter prefix can't consume it);
+		#   - blocks short names matching inside longer names (e.g.
+		#     "王昱" must not match a line "王昱涵" — the lookahead fails
+		#     because 涵 is a CJK letter).
+		m = re.search(
+			r"(?m)^[^一-鿿぀-ヿa-zA-Z]*" + re.escape(name) + r"(?![一-鿿぀-ヿa-zA-Z])",
+			ocrText,
+		)
+		if m:
+			occurrences.append((m.start(), name))
+			seen.add(name)
+	if len(occurrences) < 2:
+		return None, None
+	occurrences.sort(key=lambda t: t[0])
+	return occurrences[0][1], occurrences[1][1]
+
+
+def _findQuotedOriginal(replyIdx, quotedSender, ocrText):
+	"""Locate the original message a reply is quoting.
+
+	Replies always come AFTER the original in chat order, so we scan
+	upward from ``replyIdx`` for messages by ``quotedSender`` and pick
+	the one whose content has the largest overlap with the OCR text
+	(which contains a fragment of the quoted preview). The OCR preview
+	is shorter than the full message — substring containment in either
+	direction or longest-common-substring (≥ ``_MIN_FUZZY_OVERLAP``)
+	count as a match.
+
+	Returns ``(msg, idx)`` on success, ``(None, None)`` otherwise.
+	"""
+	if not quotedSender or replyIdx <= 0:
+		return None, None
+	ocrNorm = _normalize(ocrText)
+	if not ocrNorm:
+		return None, None
+	bestOverlap = 0
+	bestIdx = -1
+	for i in range(replyIdx - 1, -1, -1):
+		msg = _messages[i]
+		if msg.get("type") != "message":
+			continue
+		if msg.get("name") != quotedSender:
+			continue
+		contentNorm = _normalize(msg.get("content", ""))
+		if not contentNorm:
+			continue
+		if contentNorm in ocrNorm:
+			overlap = len(contentNorm)
+		elif ocrNorm in contentNorm:
+			overlap = len(ocrNorm)
+		else:
+			overlap = _longestCommonSubstring(contentNorm, ocrNorm)
+		if overlap >= _MIN_FUZZY_OVERLAP and overlap > bestOverlap:
+			bestOverlap = overlap
+			bestIdx = i
+	if bestIdx < 0:
+		return None, None
+	return _messages[bestIdx], bestIdx
+
+
+def getLastReplyInfo():
+	"""Return reply context for the most recent successful lookup.
+
+	Returns a dict with ``replySender``, ``replyContent``, ``replyTime``,
+	``originalName``, ``originalContent``, ``originalTime`` and
+	``originalIdx`` when the matched bubble was a reply, otherwise
+	``None``. ``originalContent``/``originalTime``/``originalIdx`` may
+	be ``None`` when the original message couldn't be located in the
+	cache (e.g. it was outside the exported window).
+	"""
+	return _lastReplyInfo
+
+
+def clearLastReplyInfo():
+	"""Clear cached reply context so the left-arrow handler fires only once."""
+	global _lastReplyInfo
+	_lastReplyInfo = None
+
+
 def lookupMessage(ocrText):
 	"""Return the cached message that best matches the OCR snippet.
 
@@ -268,7 +383,10 @@ def lookupMessage(ocrText):
 	Returns:
 		(formattedText, index) on match, otherwise (None, None).
 	"""
-	global _lastMatchedIdx
+	global _lastMatchedIdx, _lastReplyInfo
+	# Default to no reply context; populated below when the matched bubble
+	# turns out to be a reply.
+	_lastReplyInfo = None
 	if not isActive() or not ocrText:
 		return None, None
 
@@ -296,6 +414,13 @@ def lookupMessage(ocrText):
 	# Date group of the cursor position (−1 if no date seen yet)
 	cursorDateGroup = _messageDateGroups[cursor] if cursor < len(_messageDateGroups) else -1
 
+	# Reply pattern: when the bubble shows a quote preview, OCR contains both
+	# the actual sender's name and the quoted user's name. The quoted preview
+	# usually has a longer content overlap than the actual reply, so without
+	# this restriction the cache would return the quoted message instead of
+	# the reply that the user just navigated to.
+	replySender, quotedSender = _detectReplyNames(ocrText)
+
 	bestScore = 0.0
 	bestIdx = -1
 
@@ -307,6 +432,8 @@ def lookupMessage(ocrText):
 		msgTime = msg.get("time", "")
 		if not msgContentNorm:
 			continue
+		if replySender and msgName != replySender:
+			continue
 
 		# Try exact substring containment first — that's the strongest signal.
 		contentOverlap = 0
@@ -353,4 +480,21 @@ def lookupMessage(ocrText):
 		f" msgIdx={_messageIndexMap[bestIdx]}"
 		f" dateGroup=[{_messageDateGroups[bestIdx]}]: {_formatMessage(_messages[bestIdx])!r}",
 	)
+
+	# Reply context: when the OCR pattern indicated a reply, locate the
+	# original (quoted) message upward in the cache so the left-arrow
+	# handler can read it on demand.
+	if replySender and quotedSender:
+		matched = _messages[bestIdx]
+		originalMsg, originalIdx = _findQuotedOriginal(bestIdx, quotedSender, ocrText)
+		_lastReplyInfo = {
+			"replySender": replySender,
+			"replyContent": matched.get("content", ""),
+			"replyTime": matched.get("time", ""),
+			"originalName": quotedSender,
+			"originalContent": originalMsg.get("content", "") if originalMsg else None,
+			"originalTime": originalMsg.get("time", "") if originalMsg else None,
+			"originalIdx": originalIdx if originalMsg else None,
+		}
+
 	return _formatMessage(_messages[bestIdx]), bestIdx
diff --git a/addon/appModules/line.py b/addon/appModules/line.py
index e980c32..976d9b1 100644
--- a/addon/appModules/line.py
+++ b/addon/appModules/line.py
@@ -564,10 +564,10 @@ def _getVoiceCallConfirmationState(text):
 	if action != "join":
 		compact = _normalizeVoiceCallConfirmationLine(text)
 		lower = compact.lower()
-		hasJoinHint = any(
-			keyword in compact
-			for keyword in ("加入", "參加", "参加", "已加入", "已參加", "已参加")
-		) or "join" in lower
+		hasJoinHint = (
+			any(keyword in compact for keyword in ("加入", "參加", "参加", "已加入", "已參加", "已参加"))
+			or "join" in lower
+		)
 		if hasJoinHint and isGroup:
 			action = "join"
 
@@ -5930,12 +5930,30 @@ def _restoreClipboard(original):
 			if cacheActive:
 				cachedText, _cacheIdx = _chatCache.lookupMessage(initialOcrText)
 				if cachedText:
+					replyInfo = _chatCache.getLastReplyInfo()
+					if replyInfo:
+						# Announce as "Sender 回覆 OriginalSender Content Time"
+						# in one utterance so the reply context and the time
+						# can't be split across two announcements (the
+						# follow-up timestamp element would otherwise repeat
+						# the message with the time on its own).
+						parts = [
+							replyInfo["replySender"],
+							"回覆",
+							replyInfo["originalName"],
+							replyInfo["replyContent"],
+						]
+						if replyInfo.get("replyTime"):
+							parts.append(replyInfo["replyTime"])
+						announceText = " ".join(parts)
+					else:
+						announceText = cachedText
 					log.info(
-						f"LINE: copy-read served from chat cache: {cachedText!r}",
+						f"LINE: copy-read served from chat cache: {announceText!r}",
 					)
 					_restoreClipboard(origClip)
 					speech.cancelSpeech()
-					ui.message(cachedText)
+					ui.message(announceText)
 					return
 				else:
 					log.debug(
@@ -9403,10 +9421,36 @@ def script_navigateAndTrack(self, gesture):
 		navigating with Tab/arrows. This script sends the key through,
 		waits briefly for LINE to process it, then queries the UIA
 		focused element directly and announces it.
+
+		Special case: if the most recently announced cached message was
+		a reply, pressing left arrow speaks the original (quoted)
+		message instead of navigating, so users can hear what the reply
+		is replying to.
 		"""
 		if _suppressAddon:
 			gesture.send()
 			return
+		try:
+			keyName = gesture.mainKeyName
+		except Exception:
+			keyName = None
+		if keyName == "leftArrow":
+			try:
+				from . import _chatCache
+
+				replyInfo = _chatCache.getLastReplyInfo()
+			except Exception:
+				log.debug(
+					"LINE: chat cache reply lookup failed",
+					exc_info=True,
+				)
+				replyInfo = None
+			if replyInfo and replyInfo.get("originalContent"):
+				ui.message(
+					f"{replyInfo['originalName']} {replyInfo['originalContent']}",
+				)
+				_chatCache.clearLastReplyInfo()
+				return
 		global _lastOCRElement, _chatListMode
 		# Exiting chat list mode on Tab/Shift+Tab navigation
 		_chatListMode = False
diff --git a/tests/test_chat_cache.py b/tests/test_chat_cache.py
index 467844c..caf4fc2 100644
--- a/tests/test_chat_cache.py
+++ b/tests/test_chat_cache.py
@@ -248,6 +248,238 @@ def test_short_content_requires_time_match_to_avoid_false_positives():
 	assert "有" in formatted2
 
 
+def test_reply_bubble_matches_actual_reply_not_quoted_preview():
+	"""When a bubble shows a reply preview (sender + quoted message),
+	OCR captures both names plus the quoted text. The lookup must match
+	the actual reply content, not the longer quoted preview."""
+	_reset_cache(
+		[
+			{
+				"type": "message",
+				"name": "王昱涵",
+				"content": "然後認領想要的工作(可揚你可以休息\n1. 科系有兩個表格爆掉了要改一下\n2. 格子要合併儲存格",
+				"time": "22:20",
+			},
+			{"type": "message", "name": "陳禹安", "content": "那我用標題", "time": "10:26"},
+			{"type": "message", "name": "莊忠諺", "content": "我修結論", "time": "10:27"},
+		],
+	)
+
+	# 莊忠諺 replies to 王昱涵's earlier message. OCR captures both names
+	# plus a chunk of the quoted preview, then the actual reply "我修結論".
+	# Without the reply-sender filter, the long quoted overlap would win.
+	formatted, idx = chat_cache.lookupMessage(
+		"莊忠諺\n0 王昱涵\n然後認領想要的工作\n( 可揚你可以休息 \n我修結論",
+	)
+	assert idx == 2
+	assert "我修結論" in formatted
+
+	# Same shape for the 陳禹安 reply.
+	formatted2, idx2 = chat_cache.lookupMessage(
+		"陳禹安\n0 王昱涵\n然後認領想要的工作\n( 可揚你可以休息 \n那我用標題\n上午 10:26",
+	)
+	assert idx2 == 1
+	assert "那我用標題" in formatted2
+
+
+def test_reply_lookup_exposes_original_message_for_left_arrow():
+	"""After matching a reply, getLastReplyInfo() returns the quoted
+	original located upward in the cache, including its content for
+	the left-arrow read-aloud handler."""
+	_reset_cache(
+		[
+			{
+				"type": "message",
+				"name": "王昱涵",
+				"content": "然後認領想要的工作(可揚你可以休息\n1. 科系有兩個表格爆掉了",
+				"time": "22:20",
+			},
+			{"type": "message", "name": "陳禹安", "content": "那我用標題", "time": "10:26"},
+			{"type": "message", "name": "莊忠諺", "content": "我修結論", "time": "10:27"},
+		],
+	)
+
+	formatted, idx = chat_cache.lookupMessage(
+		"莊忠諺\n0 王昱涵\n然後認領想要的工作\n( 可揚你可以休息 \n我修結論",
+	)
+	assert idx == 2
+	assert "我修結論" in formatted
+
+	info = chat_cache.getLastReplyInfo()
+	assert info is not None
+	assert info["replySender"] == "莊忠諺"
+	assert info["replyContent"] == "我修結論"
+	assert info["originalName"] == "王昱涵"
+	assert info["originalIdx"] == 0
+	assert "然後認領想要的工作" in info["originalContent"]
+
+
+def test_reply_lookup_clears_reply_info_for_non_reply_message():
+	"""Non-reply OCR (single name) must clear stale reply info so the
+	left-arrow handler doesn't read an old original."""
+	_reset_cache(
+		[
+			{"type": "message", "name": "Alice", "content": "舊訊息", "time": "09:00"},
+			{"type": "message", "name": "Bob", "content": "回覆內容", "time": "09:05"},
+			{"type": "message", "name": "Bob", "content": "後續訊息", "time": "09:10"},
+		],
+	)
+
+	# First lookup: a reply (2 names) — populates reply info.
+	chat_cache.lookupMessage("Bob\n0 Alice\n舊訊息\n回覆內容\n上午 9 : 05")
+	assert chat_cache.getLastReplyInfo() is not None
+
+	# Next lookup: regular non-reply message — reply info must clear.
+	chat_cache.lookupMessage("Bob 後續訊息 上午 9 : 10")
+	assert chat_cache.getLastReplyInfo() is None
+
+
+def test_reply_lookup_clears_reply_info_when_no_match():
+	"""When the cache can't match anything, stale reply info must clear."""
+	_reset_cache(
+		[
+			{"type": "message", "name": "Alice", "content": "舊訊息", "time": "09:00"},
+			{"type": "message", "name": "Bob", "content": "回覆", "time": "09:05"},
+		],
+	)
+
+	chat_cache.lookupMessage("Bob\n0 Alice\n舊訊息\n回覆\n上午 9 : 05")
+	assert chat_cache.getLastReplyInfo() is not None
+
+	# Unrelated OCR — no match, reply info must clear.
+	formatted, idx = chat_cache.lookupMessage("完全不同的內容沒有時間")
+	assert formatted is None
+	assert chat_cache.getLastReplyInfo() is None
+
+
+def test_find_quoted_original_only_searches_upward():
+	"""The original message is always BEFORE the reply in chat order;
+	never match a later message even if it has the same content."""
+	_reset_cache(
+		[
+			{"type": "message", "name": "Alice", "content": "說了某句話", "time": "09:00"},
+			{"type": "message", "name": "Bob", "content": "回應", "time": "09:05"},
+			{"type": "message", "name": "Alice", "content": "說了某句話", "time": "10:00"},
+		],
+	)
+
+	# Bob (idx=1) replies to Alice's earlier message (idx=0). Even though
+	# the later Alice message (idx=2) has identical content, the search
+	# upward must pick idx=0.
+	chat_cache.lookupMessage("Bob\n0 Alice\n說了某句話\n回應\n上午 9 : 05")
+	info = chat_cache.getLastReplyInfo()
+	assert info is not None
+	assert info["originalIdx"] == 0
+
+
+def test_reply_filter_does_not_engage_when_only_one_name_in_ocr():
+	"""Single-name OCR is not a reply preview — the filter must not engage,
+	otherwise messages from anyone else become unmatchable."""
+	_reset_cache(
+		[
+			{"type": "message", "name": "Alice", "content": "早安", "time": "09:00"},
+			{"type": "message", "name": "Bob", "content": "回覆內容", "time": "09:05"},
+		],
+	)
+
+	# Only Bob appears in OCR; reply filter must not exclude Bob's own
+	# message. (If it did engage incorrectly using "Alice", Bob's message
+	# would be filtered out and we'd return None.)
+	formatted, idx = chat_cache.lookupMessage("Bob\n回覆內容\n上午 9 : 05")
+	assert idx == 1
+	assert "回覆內容" in formatted
+
+
+def test_reply_filter_not_triggered_by_name_in_message_body():
+	"""A name mentioned inside message text must not trigger reply detection.
+
+	If Alice sends '感謝Bob你的幫助' and both Alice and Bob are in the cache,
+	the old find() approach would detect two names and wrongly enter reply
+	mode.  The line-anchored regex must not match 'Bob' mid-sentence.
+	"""
+	_reset_cache(
+		[
+			{"type": "message", "name": "Bob", "content": "沒問題", "time": "09:00"},
+			{
+				"type": "message",
+				"name": "Alice",
+				"content": "感謝Bob你的幫助",
+				"time": "09:05",
+			},
+		],
+	)
+
+	# OCR for Alice's message — 'Bob' is inside the content line, not a
+	# standalone line.  lookupMessage must match Alice's message normally
+	# (idx=1) without entering reply mode.
+	formatted, idx = chat_cache.lookupMessage("Alice\n感謝Bob你的幫助\n上午 9 : 05")
+	assert idx == 1
+	assert chat_cache.getLastReplyInfo() is None
+
+
+def test_reply_detection_handles_quote_indicator_prefix():
+	"""Real LINE OCR renders a quote-indicator glyph (often "0 ") before
+	the quoted user name in reply bubbles, so the quoted name does NOT
+	occupy a standalone line.  The regex must still detect it.
+
+	Regression for an over-strict ``^name$`` regex that broke the
+	original reply-bubble fix on actual LINE OCR.
+	"""
+	_reset_cache(
+		[
+			{
+				"type": "message",
+				"name": "王昱涵",
+				"content": "然後認領想要的工作(可揚你可以休息",
+				"time": "22:20",
+			},
+			{"type": "message", "name": "莊忠諺", "content": "我修結論", "time": "10:23"},
+		],
+	)
+
+	# Exact OCR shape from the LINE log: actual sender on its own line,
+	# quoted user preceded by the "0 " quote-indicator glyph.
+	formatted, idx = chat_cache.lookupMessage(
+		"莊忠諺\n0 王昱涵\n然後認領想要的工作\n( 可揚你可以休息 \n我修結論",
+	)
+	assert idx == 1
+	assert "我修結論" in formatted
+	info = chat_cache.getLastReplyInfo()
+	assert info is not None
+	assert info["replySender"] == "莊忠諺"
+	assert info["originalName"] == "王昱涵"
+
+
+def test_reply_filter_not_triggered_by_substring_name():
+	"""A short name that is a substring of a longer name must not match.
+
+	If '王昱' and '王昱涵' are both in the cache and the OCR line is
+	'王昱涵', the regex must not match '王昱' against that line.
+	"""
+	_reset_cache(
+		[
+			{
+				"type": "message",
+				"name": "王昱",
+				"content": "你好",
+				"time": "09:00",
+			},
+			{
+				"type": "message",
+				"name": "王昱涵",
+				"content": "收到",
+				"time": "09:05",
+			},
+		],
+	)
+
+	# OCR for 王昱涵's message — only '王昱涵' occupies a standalone line.
+	# '王昱' must NOT match, so only one name is found → no reply filter.
+	formatted, idx = chat_cache.lookupMessage("王昱涵\n收到\n上午 9 : 05")
+	assert idx == 1
+	assert chat_cache.getLastReplyInfo() is None
+
+
 def test_lookup_returns_none_when_ocr_text_unrelated():
 	_reset_cache(
 		[