diff --git a/package.json b/package.json
index 234144b..da8f011 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "hermes-web-ui",
-  "version": "0.4.4",
+  "version": "0.4.5",
   "description": "Self-hosted AI chat dashboard for Hermes Agent — multi-model (Claude, GPT, Gemini, DeepSeek) web UI with Telegram, Discord, Slack, WhatsApp integration",
   "repository": {
     "type": "git",
@@ -61,13 +61,18 @@
     "dist/"
   ],
   "dependencies": {
-    "node-pty": "^1.1.0"
+    "eventsource": "^4.1.0",
+    "node-pty": "^1.1.0",
+    "socket.io": "^4.8.3",
+    "socket.io-client": "^4.8.3"
   },
   "devDependencies": {
+    "@multiavatar/multiavatar": "^1.0.7",
     "@koa/bodyparser": "^5.0.0",
     "@koa/cors": "^5.0.0",
     "@koa/router": "^15.4.0",
     "@pinia/testing": "^1.0.3",
+    "@types/eventsource": "^1.1.15",
     "@types/js-yaml": "^4.0.9",
     "@types/koa": "^2.15.0",
     "@types/koa__cors": "^5.0.0",
diff --git a/packages/client/src/api/hermes/group-chat.ts b/packages/client/src/api/hermes/group-chat.ts
new file mode 100644
index 0000000..7f8377c
--- /dev/null
+++ b/packages/client/src/api/hermes/group-chat.ts
@@ -0,0 +1,188 @@
+import { io } from 'socket.io-client'
+import { request, getApiKey } from '../client'
+
+// ─── Types ──────────────────────────────────────────────────
+
+export interface RoomInfo {
+    id: string
+    name: string
+    inviteCode: string | null
+    triggerTokens?: number
+    maxHistoryTokens?: number
+    tailMessageCount?: number
+    totalTokens?: number
+}
+
+export interface RoomAgent {
+    id: string
+    roomId: string
+    agentId: string
+    profile: string
+    name: string
+    description: string
+    invited: number
+}
+
+export interface ChatMessage {
+    id: string
+    roomId: string
+    senderId: string
+    senderName: string
+    content: string
+    timestamp: number
+}
+
+export interface MemberInfo {
+    id: string
+    userId: string
+    name: string
+    description: string
+    joinedAt: number
+}
+
+export interface JoinResult {
+    roomId: string
+    roomName: string
+    members: MemberInfo[]
+    messages: ChatMessage[]
+    rooms: string[]
+}
+
+// ─── Socket.IO Client ──────────────────────────────────────
+
+let socket: ReturnType<typeof io> | null = null
+
+export function connectGroupChat(opts?: { userId?: string; userName?: string; description?: string }): ReturnType<typeof io> {
+    if (socket?.connected) return socket
+
+    const token = getApiKey()
+    const userId = opts?.userId || localStorage.getItem('gc_user_id') || generateUUID()
+    localStorage.setItem('gc_user_id', userId)
+
+    socket = io('/group-chat', {
+        auth: {
+            token: token || undefined,
+            userId,
+            name: opts?.userName || localStorage.getItem('gc_user_name') || undefined,
+            description: opts?.description || localStorage.getItem('gc_user_description') || undefined,
+        },
+        transports: ['websocket'],
+        reconnection: true,
+        reconnectionAttempts: Infinity,
+        reconnectionDelay: 1000,
+        reconnectionDelayMax: 30000,
+    })
+
+    return socket
+}
+
+export function getStoredUserId(): string {
+    let id = localStorage.getItem('gc_user_id')
+    if (!id) {
+        id = generateUUID()
+        localStorage.setItem('gc_user_id', id)
+    }
+    return id
+}
+
+export function getStoredUserName(): string | null {
+    return localStorage.getItem('gc_user_name')
+}
+
+function generateUUID(): string {
+    return 'xxxxxxxx-xxxx-4xxx-yxxx-xxxxxxxxxxxx'.replace(/[xy]/g, (c) => {
+        const r = Math.random() * 16 | 0
+        const v = c === 'x' ? r : (r & 0x3 | 0x8)
+        return v.toString(16)
+    })
+}
+
+export function getSocket(): ReturnType<typeof io> | null {
+    return socket?.connected ? socket : null
+}
+
+export function disconnectGroupChat(): void {
+    if (socket) {
+        socket.disconnect()
+        socket = null
+    }
+}
+
+// ─── REST API ───────────────────────────────────────────────
+
+export async function createRoom(data: {
+    name: string
+    inviteCode: string
+    agents?: { profile: string; name?: string; description?: string; invited?: boolean }[]
+    compression?: { triggerTokens?: number; maxHistoryTokens?: number; tailMessageCount?: number }
+}): Promise<{ room: RoomInfo; agents: RoomAgent[] }> {
+    return request('/api/hermes/group-chat/rooms', {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(data),
+    })
+}
+
+export async function listRooms(): Promise<{ rooms: RoomInfo[] }> {
+    return request('/api/hermes/group-chat/rooms')
+}
+
+export async function getRoomDetail(roomId: string): Promise<{ room: RoomInfo; messages: ChatMessage[]; agents: RoomAgent[]; members: MemberInfo[] }> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}`)
+}
+
+export async function joinRoomByCode(code: string): Promise<{ room: RoomInfo }> {
+    return request(`/api/hermes/group-chat/rooms/join/${code}`)
+}
+
+export async function updateInviteCode(roomId: string, inviteCode: string): Promise<void> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}/invite-code`, {
+        method: 'PUT',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ inviteCode }),
+    })
+}
+
+export async function addAgent(roomId: string, data: {
+    profile: string
+    name?: string
+    description?: string
+    invited?: boolean
+}): Promise<{ agent: RoomAgent }> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}/agents`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(data),
+    })
+}
+
+export async function listAgents(roomId: string): Promise<{ agents: RoomAgent[] }> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}/agents`)
+}
+
+export async function removeAgent(roomId: string, agentId: string): Promise<void> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}/agents/${agentId}`, {
+        method: 'DELETE',
+    })
+}
+
+export async function deleteRoom(roomId: string): Promise<void> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}`, {
+        method: 'DELETE',
+    })
+}
+
+export async function updateRoomConfig(roomId: string, config: { triggerTokens?: number; maxHistoryTokens?: number; tailMessageCount?: number }): Promise<{ room: RoomInfo }> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}/config`, {
+        method: 'PUT',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify(config),
+    })
+}
+
+export async function forceCompress(roomId: string): Promise<{ success: boolean; summary: string }> {
+    return request(`/api/hermes/group-chat/rooms/${roomId}/compress`, {
+        method: 'POST',
+    })
+}
+
diff --git a/packages/client/src/components/hermes/chat/MarkdownRenderer.vue b/packages/client/src/components/hermes/chat/MarkdownRenderer.vue
index a04f441..5c62604 100644
--- a/packages/client/src/components/hermes/chat/MarkdownRenderer.vue
+++ b/packages/client/src/components/hermes/chat/MarkdownRenderer.vue
@@ -6,7 +6,13 @@ import MarkdownIt from 'markdown-it'
 import { handleCodeBlockCopyClick, renderHighlightedCodeBlock } from './highlight'
 import { downloadFile } from '@/api/hermes/download'
 
-const props = defineProps<{ content: string }>()
+const props = withDefaults(defineProps<{
+    content: string
+    mentionNames?: string[]
+}>(), {
+    mentionNames: () => [],
+})
+
 const { t } = useI18n()
 const message = useMessage()
 
@@ -19,7 +25,15 @@ const md: MarkdownIt = new MarkdownIt({
   },
 })
 
-const renderedHtml = computed(() => md.render(props.content))
+const renderedHtml = computed(() => {
+  let html = md.render(props.content)
+  if (props.mentionNames && props.mentionNames.length > 0) {
+    const escaped = props.mentionNames.map(n => n.replace(/[.*+?^${}()|[\]\\]/g, '\\$&'))
+    const re = new RegExp(`(?<=[\\s>]|^)@(${escaped.join('|')})(?=\\s|$)`, 'gi')
+    html = html.replace(re, '<span class="mention-highlight">@$1</span>')
+  }
+  return html
+})
 
 function handleMarkdownClick(event: MouseEvent): void {
   void handleCodeBlockCopyClick(event)
diff --git a/packages/client/src/components/hermes/group-chat/CreateRoomForm.vue b/packages/client/src/components/hermes/group-chat/CreateRoomForm.vue
new file mode 100644
index 0000000..7a52703
--- /dev/null
+++ b/packages/client/src/components/hermes/group-chat/CreateRoomForm.vue
@@ -0,0 +1,160 @@
+<script setup lang="ts">
+import { ref, nextTick } from 'vue'
+import { useI18n } from 'vue-i18n'
+import { NInput, NButton, NSpace, NInputNumber, NCollapse, NCollapseItem } from 'naive-ui'
+
+type InputLikeInstance = {
+    focus: () => void
+}
+
+const { t } = useI18n()
+const emit = defineEmits<{
+    submit: [name: string, inviteCode: string, userName: string, description: string, compression: { triggerTokens: number; maxHistoryTokens: number; tailMessageCount: number }]
+    cancel: []
+}>()
+
+const roomName = ref('')
+const inviteCode = ref('')
+const userName = ref('')
+const description = ref('')
+const roomInput = ref<InputLikeInstance | null>(null)
+
+const compression = ref({
+    triggerTokens: 100000,
+    maxHistoryTokens: 32000,
+    tailMessageCount: 20,
+})
+
+function generateCode(): string {
+    const chars = 'ABCDEFGHJKLMNPQRSTUVWXYZ23456789'
+    let code = ''
+    for (let i = 0; i < 6; i++) {
+        code += chars[Math.floor(Math.random() * chars.length)]
+    }
+    return code
+}
+
+function handleCreate() {
+    const name = roomName.value.trim()
+    const code = inviteCode.value.trim() || generateCode()
+    const user = userName.value.trim()
+    if (!name || !user) return
+    emit('submit', name, code, user, description.value.trim(), { ...compression.value })
+}
+
+function focusRoomInput() {
+    nextTick(() => roomInput.value?.focus())
+}
+</script>
+
+<template>
+    <div class="create-form">
+        <div class="form-group">
+            <label class="form-label">{{ t('groupChat.yourName') }}</label>
+            <NInput
+                v-model:value="userName"
+                :placeholder="t('groupChat.yourNamePlaceholder')"
+                @keyup.enter="focusRoomInput"
+            />
+        </div>
+        <div class="form-group">
+            <label class="form-label">{{ t('groupChat.yourDescription') }}</label>
+            <NInput
+                v-model:value="description"
+                type="textarea"
+                :rows="2"
+                :placeholder="t('groupChat.yourDescriptionPlaceholder')"
+            />
+        </div>
+        <div class="form-group">
+            <label class="form-label">{{ t('groupChat.roomName') }}</label>
+            <NInput
+                ref="roomInput"
+                v-model:value="roomName"
+                :placeholder="t('groupChat.roomNamePlaceholder')"
+                @keyup.enter="handleCreate"
+            />
+        </div>
+        <div class="form-group">
+            <label class="form-label">{{ t('groupChat.inviteCode') }}</label>
+            <div class="code-row">
+                <NInput
+                    v-model:value="inviteCode"
+                    :placeholder="t('groupChat.autoGenerate')"
+                />
+                <NButton size="small" @click="inviteCode = generateCode()">
+                    <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
+                        <polyline points="23 4 23 10 17 10" /><path d="M20.49 15a9 9 0 1 1-2.12-9.36L23 10" />
+                    </svg>
+                </NButton>
+            </div>
+        </div>
+
+        <NCollapse class="compression-collapse">
+            <NCollapseItem :title="t('groupChat.compressionSettings')" name="compression">
+                <div class="compression-fields">
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.triggerTokens') }}</label>
+                        <NInputNumber v-model:value="compression.triggerTokens" :min="1000" :step="1000" style="width: 100%" />
+                        <p class="form-hint">{{ t('groupChat.triggerTokensDesc') }}</p>
+                    </div>
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.maxHistoryTokens') }}</label>
+                        <NInputNumber v-model:value="compression.maxHistoryTokens" :min="1000" :step="1000" style="width: 100%" />
+                        <p class="form-hint">{{ t('groupChat.maxHistoryTokensDesc') }}</p>
+                    </div>
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.tailMessageCount') }}</label>
+                        <NInputNumber v-model:value="compression.tailMessageCount" :min="1" :step="5" style="width: 100%" />
+                        <p class="form-hint">{{ t('groupChat.tailMessageCountDesc') }}</p>
+                    </div>
+                </div>
+            </NCollapseItem>
+        </NCollapse>
+
+        <div class="modal-actions">
+            <NSpace justify="end">
+                <NButton @click="emit('cancel')">{{ t('common.cancel') }}</NButton>
+                <NButton type="primary" :disabled="!roomName.trim() || !userName.trim()" @click="handleCreate">{{ t('common.create') }}</NButton>
+            </NSpace>
+        </div>
+    </div>
+</template>
+
+<style scoped lang="scss">
+@use "@/styles/variables" as *;
+
+.create-form {
+    .form-group {
+        margin-bottom: 16px;
+    }
+}
+
+.form-label {
+    display: block;
+    font-size: 13px;
+    font-weight: 500;
+    color: $text-secondary;
+    margin-bottom: 6px;
+}
+
+.form-hint {
+    font-size: 11px;
+    color: $text-muted;
+    margin: 4px 0 0;
+}
+
+.code-row {
+    display: flex;
+    gap: 8px;
+    align-items: center;
+}
+
+.compression-collapse {
+    margin-bottom: 16px;
+}
+
+.compression-fields {
+    padding-top: 8px;
+}
+</style>
diff --git a/packages/client/src/components/hermes/group-chat/GroupChatInput.vue b/packages/client/src/components/hermes/group-chat/GroupChatInput.vue
new file mode 100644
index 0000000..684c7ae
--- /dev/null
+++ b/packages/client/src/components/hermes/group-chat/GroupChatInput.vue
@@ -0,0 +1,321 @@
+<script setup lang="ts">
+import { ref, computed, nextTick } from 'vue'
+import { useI18n } from 'vue-i18n'
+import { NButton, NDropdown } from 'naive-ui'
+import { useGroupChatStore } from '@/stores/hermes/group-chat'
+import type { DropdownOption } from 'naive-ui'
+
+const { t } = useI18n()
+const emit = defineEmits<{ send: [content: string] }>()
+const store = useGroupChatStore()
+
+const inputText = ref('')
+const textareaRef = ref<HTMLTextAreaElement>()
+const isComposing = ref(false)
+
+// ─── Mention State ───────────────────────────────────────
+
+const mentionActive = ref(false)
+const mentionQuery = ref('')
+const mentionStartIndex = ref(-1)
+const dropdownX = ref(0)
+const dropdownY = ref(0)
+const activeIndex = ref(0)
+
+const filteredAgents = computed(() => {
+    const query = mentionQuery.value.toLowerCase()
+    return store.agents.filter(a => a.name.toLowerCase().includes(query))
+})
+
+const dropdownOptions = computed<DropdownOption[]>(() => {
+    return filteredAgents.value.map((a, i) => ({
+        label: `${a.name}  (${a.profile})`,
+        key: a.name,
+        index: i,
+    }))
+})
+
+const canSend = computed(() => !!inputText.value.trim())
+
+// ─── Mention Logic ───────────────────────────────────────
+
+function updateMentionState() {
+    const el = textareaRef.value
+    if (!el) { mentionActive.value = false; return }
+
+    const text = inputText.value
+    const cursorPos = el.selectionStart
+
+    // Find the last @ before the cursor
+    let atPos = -1
+    for (let i = cursorPos - 1; i >= 0; i--) {
+        if (text[i] === '@') { atPos = i; break }
+        if (text[i] === ' ' || text[i] === '\n') break
+    }
+
+    if (atPos === -1) {
+        mentionActive.value = false
+        return
+    }
+
+    // Make sure the @ is not part of a word (preceded by space or start of line)
+    if (atPos > 0 && text[atPos - 1] !== ' ' && text[atPos - 1] !== '\n') {
+        mentionActive.value = false
+        return
+    }
+
+    const query = text.slice(atPos + 1, cursorPos)
+    if (query.includes(' ')) {
+        mentionActive.value = false
+        return
+    }
+
+    mentionQuery.value = query
+    mentionStartIndex.value = atPos
+    activeIndex.value = 0
+
+    // Calculate dropdown position using mirror span
+    const mirror = document.createElement('span')
+    const style = getComputedStyle(el)
+    const props = ['fontFamily', 'fontSize', 'fontWeight', 'letterSpacing', 'textTransform', 'wordSpacing', 'textIndent', 'border', 'padding', 'boxSizing', 'lineHeight']
+    props.forEach(p => { (mirror.style as any)[p] = style[p as any] })
+    mirror.style.position = 'absolute'
+    mirror.style.visibility = 'hidden'
+    mirror.style.whiteSpace = 'nowrap'
+    mirror.textContent = text.slice(0, atPos + 1)
+
+    const rect = el.getBoundingClientRect()
+    document.body.appendChild(mirror)
+    const mirrorRect = mirror.getBoundingClientRect()
+    document.body.removeChild(mirror)
+
+    dropdownX.value = rect.left + mirrorRect.width - el.scrollLeft
+    // Estimate Y based on newlines before @ and line height
+    const linesBeforeAt = text.slice(0, atPos + 1).split('\n').length - 1
+    const lineHeight = parseFloat(style.lineHeight) || parseFloat(style.fontSize) * 1.2
+    dropdownY.value = rect.top + linesBeforeAt * lineHeight - el.scrollTop + lineHeight + 4
+
+    mentionActive.value = filteredAgents.value.length > 0
+}
+
+function selectMention(name: string) {
+    const el = textareaRef.value
+    if (!el || mentionStartIndex.value === -1) return
+
+    const before = inputText.value.slice(0, mentionStartIndex.value)
+    const after = inputText.value.slice(el.selectionStart)
+    inputText.value = `${before}@${name} ${after}`
+    mentionActive.value = false
+
+    nextTick(() => {
+        if (el) {
+            const newPos = before.length + name.length + 2
+            el.setSelectionRange(newPos, newPos)
+            el.focus()
+            el.style.height = 'auto'
+            el.style.height = Math.min(el.scrollHeight, 100) + 'px'
+        }
+    })
+}
+
+// ─── Event Handlers ──────────────────────────────────────
+
+function handleKeydown(e: KeyboardEvent) {
+    // Mention navigation
+    if (mentionActive.value && filteredAgents.value.length > 0) {
+        if (e.key === 'ArrowDown') {
+            e.preventDefault()
+            activeIndex.value = (activeIndex.value + 1) % filteredAgents.value.length
+            return
+        }
+        if (e.key === 'ArrowUp') {
+            e.preventDefault()
+            activeIndex.value = (activeIndex.value - 1 + filteredAgents.value.length) % filteredAgents.value.length
+            return
+        }
+        if (e.key === 'Enter' || e.key === 'Tab') {
+            e.preventDefault()
+            selectMention(filteredAgents.value[activeIndex.value].name)
+            return
+        }
+        if (e.key === 'Escape') {
+            e.preventDefault()
+            mentionActive.value = false
+            return
+        }
+    }
+
+    if (e.key !== 'Enter' || e.shiftKey) return
+    if (isComposing.value || e.isComposing || e.keyCode === 229) return
+    e.preventDefault()
+    handleSend()
+}
+
+function handleSend() {
+    const content = inputText.value.trim()
+    if (!content) return
+
+    emit('send', content)
+    inputText.value = ''
+    mentionActive.value = false
+
+    nextTick(() => {
+        if (textareaRef.value) {
+            textareaRef.value.style.height = 'auto'
+        }
+    })
+}
+
+function handleInput(e: Event) {
+    store.emitTyping()
+    const el = e.target as HTMLTextAreaElement
+    el.style.height = 'auto'
+    el.style.height = Math.min(el.scrollHeight, 100) + 'px'
+
+    if (!isComposing.value) {
+        updateMentionState()
+    }
+}
+
+function handleDropdownSelect(key: string) {
+    selectMention(key)
+}
+
+function handleDropdownClickOutside() {
+    mentionActive.value = false
+}
+
+function handleCompositionStart() {
+    isComposing.value = true
+}
+
+function handleCompositionEnd() {
+    requestAnimationFrame(() => {
+        isComposing.value = false
+        updateMentionState()
+    })
+}
+</script>
+
+<template>
+    <div class="chat-input-area">
+        <div class="input-wrapper">
+            <textarea
+                ref="textareaRef"
+                v-model="inputText"
+                class="input-textarea"
+                :placeholder="t('groupChat.inputPlaceholder')"
+                rows="1"
+                @keydown="handleKeydown"
+                @compositionstart="handleCompositionStart"
+                @compositionend="handleCompositionEnd"
+                @input="handleInput"
+            />
+            <div class="input-actions">
+                <NButton
+                    size="small"
+                    type="primary"
+                    :disabled="!canSend"
+                    @click="handleSend"
+                >
+                    <template #icon>
+                        <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><line x1="22" y1="2" x2="11" y2="13"/><polygon points="22 2 15 22 11 13 2 9 22 2"/></svg>
+                    </template>
+                    {{ t('chat.send') }}
+                </NButton>
+            </div>
+        </div>
+        <NDropdown
+            placement="top-start"
+            trigger="manual"
+            :x="dropdownX"
+            :y="dropdownY"
+            :options="dropdownOptions"
+            :show="mentionActive"
+            :render-icon="() => null"
+            @select="handleDropdownSelect"
+            @clickoutside="handleDropdownClickOutside"
+        />
+    </div>
+</template>
+
+<style scoped lang="scss">
+@use "@/styles/variables" as *;
+
+.chat-input-area {
+    padding: 20px 20px 16px;
+    border-top: 1px solid $border-color;
+    flex-shrink: 0;
+}
+
+.typing-dots {
+    display: inline-flex;
+    align-items: center;
+    gap: 2px;
+
+    span {
+        display: block;
+        width: 4px;
+        height: 4px;
+        border-radius: 50%;
+        background-color: $text-muted;
+        animation: typing-bounce 1.2s infinite;
+
+        &:nth-child(2) { animation-delay: 0.2s; }
+        &:nth-child(3) { animation-delay: 0.4s; }
+    }
+}
+
+@keyframes typing-bounce {
+    0%, 60%, 100% { transform: translateY(0); opacity: 0.4; }
+    30% { transform: translateY(-3px); opacity: 1; }
+}
+
+.input-wrapper {
+    display: flex;
+    align-items: center;
+    gap: 10px;
+    background-color: $bg-input;
+    border: 1px solid $border-color;
+    border-radius: $radius-md;
+    padding: 10px 12px;
+    transition: border-color $transition-fast, background-color $transition-fast;
+
+    &:focus-within {
+        border-color: $accent-primary;
+    }
+
+    .dark & {
+        background-color: #333333;
+    }
+}
+
+.input-textarea {
+    flex: 1;
+    background: none;
+    border: none;
+    outline: none;
+    color: $text-primary;
+    font-family: $font-ui;
+    font-size: 14px;
+    line-height: 1.5;
+    resize: none;
+    max-height: 100px;
+    min-height: 20px;
+    overflow-y: auto;
+
+    &::placeholder {
+        color: $text-muted;
+        white-space: nowrap;
+        overflow: hidden;
+        text-overflow: ellipsis;
+    }
+}
+
+.input-actions {
+    display: flex;
+    gap: 6px;
+    flex-shrink: 0;
+    align-items: center;
+}
+</style>
diff --git a/packages/client/src/components/hermes/group-chat/GroupChatPanel.vue b/packages/client/src/components/hermes/group-chat/GroupChatPanel.vue
new file mode 100644
index 0000000..68c5f9f
--- /dev/null
+++ b/packages/client/src/components/hermes/group-chat/GroupChatPanel.vue
@@ -0,0 +1,896 @@
+<script setup lang="ts">
+import { ref, computed, nextTick, watch } from 'vue'
+import { useI18n } from 'vue-i18n'
+import { useMessage, NInput, NButton, NSpace, NSelect, NPopover, NPopconfirm, NInputNumber } from 'naive-ui'
+import multiavatar from '@multiavatar/multiavatar'
+import { useGroupChatStore } from '@/stores/hermes/group-chat'
+import { useProfilesStore } from '@/stores/hermes/profiles'
+import { updateRoomConfig, forceCompress } from '@/api/hermes/group-chat'
+import GroupMessageList from './GroupMessageList.vue'
+import GroupChatInput from './GroupChatInput.vue'
+
+const { t } = useI18n()
+const message = useMessage()
+const store = useGroupChatStore()
+const profilesStore = useProfilesStore()
+
+const showSidebar = ref(true)
+const showCreateModal = ref(false)
+const showAddAgentModal = ref(false)
+const showCompressionModal = ref(false)
+const compressionConfig = ref({ triggerTokens: 100000, maxHistoryTokens: 32000, tailMessageCount: 20 })
+const isCompressing = ref(false)
+const selectedProfile = ref<string | null>(null)
+const agentName = ref('')
+const agentDescription = ref('')
+
+const profileOptions = computed(() =>
+    profilesStore.profiles.map(p => ({ label: p.name, value: p.name }))
+)
+
+const avatarCache = new Map<string, string>()
+
+function agentAvatarUrl(name: string): string {
+    if (avatarCache.has(name)) return avatarCache.get(name)!
+    const uri = multiavatar(name)
+    avatarCache.set(name, uri)
+    return uri
+}
+
+const hasRoom = computed(() => !!store.currentRoomId)
+
+function formatTokens(tokens: number): string {
+    if (tokens >= 1000) return `${(tokens / 1000).toFixed(1)}k tokens`
+    return `${tokens} tokens`
+}
+
+function toggleSidebar() {
+    showSidebar.value = !showSidebar.value
+}
+
+async function handleCreateRoom(name: string, inviteCode: string, userName: string, description: string, compression: { triggerTokens: number; maxHistoryTokens: number; tailMessageCount: number }) {
+    try {
+        store.setUserInfo(userName, description)
+        const res = await store.createNewRoom(name, inviteCode, undefined, compression)
+        showCreateModal.value = false
+        message.success(t('groupChat.roomCreated'))
+        await store.joinRoom(res.room.id)
+    } catch {
+        message.error(t('common.saveFailed'))
+    }
+}
+
+async function handleDeleteRoom(roomId: string) {
+    try {
+        await store.deleteRoom(roomId)
+        message.success(t('groupChat.roomDeleted'))
+    } catch {
+        message.error(t('common.saveFailed'))
+    }
+}
+
+async function handleSelectRoom(roomId: string) {
+    try {
+        await store.joinRoom(roomId)
+    } catch {
+        message.error(t('groupChat.joinFailed'))
+    }
+}
+
+async function handleSendMessage(content: string) {
+    try {
+        await store.sendMessage(content)
+    } catch (err: any) {
+        message.error(err.message)
+    }
+}
+
+async function handleAddAgent() {
+    await profilesStore.fetchProfiles()
+    showAddAgentModal.value = true
+}
+
+async function confirmAddAgent() {
+    if (!selectedProfile.value || !store.currentRoomId) return
+    try {
+        await store.addAgentToRoom(store.currentRoomId, {
+            profile: selectedProfile.value,
+            name: agentName.value.trim() || undefined,
+            description: agentDescription.value.trim() || undefined,
+        })
+        showAddAgentModal.value = false
+        selectedProfile.value = null
+        agentName.value = ''
+        agentDescription.value = ''
+        message.success(t('groupChat.agentAdded'))
+    } catch (err: any) {
+        if (err.message?.includes('already')) {
+            message.warning(t('groupChat.agentAlreadyInRoom'))
+        } else {
+            message.error(t('common.saveFailed'))
+        }
+    }
+}
+
+function handleOpenCompressionConfig() {
+    const room = store.rooms.find(r => r.id === store.currentRoomId)
+    if (room) {
+        compressionConfig.value = {
+            triggerTokens: room.triggerTokens ?? 100000,
+            maxHistoryTokens: room.maxHistoryTokens ?? 32000,
+            tailMessageCount: room.tailMessageCount ?? 20,
+        }
+    }
+    showCompressionModal.value = true
+}
+
+async function handleSaveCompressionConfig() {
+    if (!store.currentRoomId) return
+    try {
+        const res = await updateRoomConfig(store.currentRoomId, { ...compressionConfig.value })
+        const idx = store.rooms.findIndex(r => r.id === store.currentRoomId)
+        if (idx >= 0 && res.room) store.rooms[idx] = res.room
+        showCompressionModal.value = false
+        message.success(t('groupChat.compressionSaved'))
+    } catch {
+        message.error(t('common.saveFailed'))
+    }
+}
+
+async function handleForceCompress() {
+    if (!store.currentRoomId || isCompressing.value) return
+    if (store.contextStatuses.size > 0) {
+        message.warning(t('groupChat.compressingInProgress'))
+        return
+    }
+    isCompressing.value = true
+    try {
+        await forceCompress(store.currentRoomId)
+        message.success(t('groupChat.compressionSaved'))
+    } catch {
+        message.error(t('common.saveFailed'))
+    } finally {
+        isCompressing.value = false
+    }
+}
+
+async function handleRemoveAgent(agentId: string) {
+    if (!store.currentRoomId) return
+    try {
+        await store.removeAgentFromRoom(store.currentRoomId, agentId)
+    } catch {
+        message.error(t('common.deleteFailed'))
+    }
+}
+
+// Auto-scroll on new messages
+const messageListRef = ref()
+watch(() => store.sortedMessages.length, async () => {
+    await nextTick()
+    messageListRef.value?.scrollToBottom()
+})
+</script>
+
+<template>
+    <div class="group-chat-panel">
+        <!-- Room sidebar -->
+        <div v-if="showSidebar" class="room-sidebar">
+            <div class="sidebar-header">
+                <span class="sidebar-title">{{ t('groupChat.title') }}</span>
+                <div class="sidebar-actions">
+                    <button class="icon-btn" :title="t('groupChat.createRoom')" @click="showCreateModal = true">
+                        <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
+                            <line x1="12" y1="5" x2="12" y2="19" /><line x1="5" y1="12" x2="19" y2="12" />
+                        </svg>
+                    </button>
+                </div>
+            </div>
+            <div class="room-list">
+                <div
+                    v-for="room in store.rooms"
+                    :key="room.id"
+                    class="room-item"
+                    :class="{ active: store.currentRoomId === room.id }"
+                    @click="handleSelectRoom(room.id)"
+                >
+                    <svg class="room-icon" width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
+                        <path d="M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z" />
+                    </svg>
+                    <div class="room-info">
+                        <span class="room-name">{{ room.name || room.id }}</span>
+                        <span v-if="room.inviteCode" class="room-code">{{ room.inviteCode }}</span>
+                        <span class="room-tokens">{{ formatTokens(room.totalTokens || 0) }}</span>
+                    </div>
+                    <NPopconfirm @positive-click="handleDeleteRoom(room.id)">
+                        <template #trigger>
+                            <button class="room-delete-btn" @click.stop>
+                                <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg>
+                            </button>
+                        </template>
+                        {{ t('groupChat.deleteRoomConfirm') }}
+                    </NPopconfirm>
+                </div>
+                <div v-if="store.rooms.length === 0" class="empty-rooms">
+                    {{ t('groupChat.noRooms') }}
+                </div>
+            </div>
+        </div>
+
+        <!-- Main chat area -->
+        <div class="chat-main">
+            <div class="chat-header">
+                <button class="icon-btn" @click="toggleSidebar">
+                    <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
+                        <rect x="3" y="3" width="18" height="18" rx="2" ry="2" /><line x1="9" y1="3" x2="9" y2="21" />
+                    </svg>
+                </button>
+                <span class="room-title-text">{{ store.roomName || (store.currentRoomId || t('groupChat.title')) }}</span>
+                <div class="header-info">
+                    <!-- Stacked avatars (user + agents) -->
+                    <NPopover v-if="store.agents.length" trigger="click" placement="bottom-end" :width="220">
+                        <template #trigger>
+                            <div class="avatar-stack-inner">
+                                <!-- User avatar first -->
+                                <span class="avatar-stack-item" :style="{ zIndex: store.agents.length + 1 }">
+                                    <span class="agent-avatar" v-html="agentAvatarUrl(store.userName || store.userId)" />
+                                </span>
+                                <span
+                                    v-for="(agent, index) in store.agents.slice(-4)"
+                                    :key="agent.id"
+                                    class="avatar-stack-item"
+                                    :style="{ zIndex: store.agents.length - index }"
+                                >
+                                    <span class="agent-avatar" v-html="agentAvatarUrl(agent.name)" />
+                                </span>
+                                <span v-if="store.agents.length > 4" class="avatar-stack-more">+{{ store.agents.length - 4 }}</span>
+                            </div>
+                        </template>
+                        <div class="agent-popover">
+                            <div class="agent-popover-item" style="margin-bottom: 8px; padding-bottom: 8px; border-bottom: 1px solid var(--n-border-color, #efeff5);">
+                                <span class="agent-avatar" v-html="agentAvatarUrl(store.userName || store.userId)" />
+                                <div class="agent-popover-info">
+                                    <span class="agent-popover-name">{{ store.userName || 'You' }}</span>
+                                    <span class="agent-popover-profile">{{ t('groupChat.you') }}</span>
+                                </div>
+                            </div>
+                            <div class="agent-popover-title">{{ t('groupChat.agents') }} ({{ store.agents.length }})</div>
+                            <div v-for="agent in store.agents" :key="agent.id" class="agent-popover-item">
+                                <span class="agent-avatar" v-html="agentAvatarUrl(agent.name)" />
+                                <div class="agent-popover-info">
+                                    <span class="agent-popover-name">{{ agent.name }}</span>
+                                    <span class="agent-popover-profile">{{ agent.profile }}</span>
+                                </div>
+                                <button class="agent-popover-remove" @click="handleRemoveAgent(agent.id)">
+                                    <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><line x1="18" y1="6" x2="6" y2="18"/><line x1="6" y1="6" x2="18" y2="18"/></svg>
+                                </button>
+                            </div>
+                        </div>
+                    </NPopover>
+                    <!-- Only user avatar, no agents -->
+                    <div v-else-if="store.userName" class="avatar-stack-inner">
+                        <span class="avatar-stack-item">
+                            <span class="agent-avatar" v-html="agentAvatarUrl(store.userName || store.userId)" />
+                        </span>
+                    </div>
+                    <button class="icon-btn" :title="t('groupChat.addAgent')" @click="handleAddAgent">
+                        <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5"><line x1="12" y1="5" x2="12" y2="19"/><line x1="5" y1="12" x2="19" y2="12"/></svg>
+                    </button>
+                    <button class="icon-btn" :title="t('groupChat.compressionConfig')" @click="handleOpenCompressionConfig">
+                        <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5"><circle cx="12" cy="12" r="3"/><path d="M19.4 15a1.65 1.65 0 0 0 .33 1.82l.06.06a2 2 0 0 1 0 2.83 2 2 0 0 1-2.83 0l-.06-.06a1.65 1.65 0 0 0-1.82-.33 1.65 1.65 0 0 0-1 1.51V21a2 2 0 0 1-2 2 2 2 0 0 1-2-2v-.09A1.65 1.65 0 0 0 9 19.4a1.65 1.65 0 0 0-1.82.33l-.06.06a2 2 0 0 1-2.83 0 2 2 0 0 1 0-2.83l.06-.06A1.65 1.65 0 0 0 4.68 15a1.65 1.65 0 0 0-1.51-1H3a2 2 0 0 1-2-2 2 2 0 0 1 2-2h.09A1.65 1.65 0 0 0 4.6 9a1.65 1.65 0 0 0-.33-1.82l-.06-.06a2 2 0 0 1 0-2.83 2 2 0 0 1 2.83 0l.06.06A1.65 1.65 0 0 0 9 4.68a1.65 1.65 0 0 0 1-1.51V3a2 2 0 0 1 2-2 2 2 0 0 1 2 2v.09a1.65 1.65 0 0 0 1 1.51 1.65 1.65 0 0 0 1.82-.33l.06-.06a2 2 0 0 1 2.83 0 2 2 0 0 1 0 2.83l-.06.06A1.65 1.65 0 0 0 19.4 4.6a1.65 1.65 0 0 0 1.51 1V3a2 2 0 0 1 2-2 2 2 0 0 1 2 2v.09a1.65 1.65 0 0 0 1.51 1z"/></svg>
+                    </button>
+                    <span v-if="store.members.length" class="member-count">
+                        {{ store.members.length }} {{ t('groupChat.members') }}
+                    </span>
+                    <span class="connection-dot" :class="{ connected: store.connected, disconnected: !store.connected }"></span>
+                </div>
+            </div>
+
+            <template v-if="hasRoom">
+                <GroupMessageList ref="messageListRef" />
+                <div v-if="store.contextStatuses.size > 0 || (store.typingText && store.contextStatuses.size === 0)" class="status-bar">
+                    <div v-if="store.contextStatuses.size > 0" class="context-status-list">
+                        <div v-for="[name, status] in store.contextStatuses" :key="name" class="context-status">
+                            <span class="typing-dots">
+                                <span /><span /><span />
+                            </span>
+                            <span v-if="status.status === 'compressing'">
+                                @{{ status.agentName }} {{ t('groupChat.agentCompressing') }}
+                            </span>
+                            <span v-else>
+                                @{{ status.agentName }} {{ t('groupChat.agentReplying') }}
+                            </span>
+                        </div>
+                    </div>
+                    <div v-else-if="store.typingText" class="typing-indicator">
+                        <span class="typing-dots">
+                            <span /><span /><span />
+                        </span>
+                        {{ store.typingText }}
+                    </div>
+                </div>
+                <GroupChatInput @send="handleSendMessage" />
+            </template>
+
+            <div v-else class="no-room">
+                <div class="no-room-icon">
+                    <svg width="48" height="48" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1" stroke-linecap="round" stroke-linejoin="round">
+                        <path d="M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z" />
+                    </svg>
+                </div>
+                <p>{{ t('groupChat.selectOrCreate') }}</p>
+            </div>
+        </div>
+
+        <!-- Create room modal -->
+        <Teleport to="body">
+            <div v-if="showCreateModal" class="modal-backdrop" @click.self="showCreateModal = false">
+                <div class="modal">
+                    <h3>{{ t('groupChat.createRoom') }}</h3>
+                    <CreateRoomForm @submit="handleCreateRoom" @cancel="showCreateModal = false" />
+                </div>
+            </div>
+        </Teleport>
+
+        <Teleport to="body">
+            <div v-if="showAddAgentModal" class="modal-backdrop" @click.self="showAddAgentModal = false">
+                <div class="modal">
+                    <h3>{{ t('groupChat.addAgent') }}</h3>
+                    <div class="form-group">
+                        <NSelect
+                            v-model:value="selectedProfile"
+                            :options="profileOptions"
+                            :placeholder="t('groupChat.selectProfile')"
+                            filterable
+                        />
+                    </div>
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.agentName') }}</label>
+                        <NInput
+                            v-model:value="agentName"
+                            :placeholder="t('groupChat.agentNamePlaceholder')"
+                        />
+                    </div>
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.agentDesc') }}</label>
+                        <NInput
+                            v-model:value="agentDescription"
+                            type="textarea"
+                            :rows="2"
+                            :placeholder="t('groupChat.agentDescPlaceholder')"
+                        />
+                    </div>
+                    <div class="modal-actions">
+                        <NSpace justify="end">
+                            <NButton @click="showAddAgentModal = false">{{ t('common.cancel') }}</NButton>
+                            <NButton type="primary" :disabled="!selectedProfile" @click="confirmAddAgent">{{ t('common.add') }}</NButton>
+                        </NSpace>
+                    </div>
+                </div>
+            </div>
+            <div v-if="showCompressionModal" class="modal-backdrop" @click.self="showCompressionModal = false">
+                <div class="modal">
+                    <h3>{{ t('groupChat.compressionConfig') }}</h3>
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.triggerTokens') }}</label>
+                        <NInputNumber v-model:value="compressionConfig.triggerTokens" :min="1000" :step="10000" style="width: 100%" />
+                        <p class="form-hint">{{ t('groupChat.triggerTokensDesc') }}</p>
+                    </div>
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.maxHistoryTokens') }}</label>
+                        <NInputNumber v-model:value="compressionConfig.maxHistoryTokens" :min="1000" :step="1000" style="width: 100%" />
+                        <p class="form-hint">{{ t('groupChat.maxHistoryTokensDesc') }}</p>
+                    </div>
+                    <div class="form-group">
+                        <label class="form-label">{{ t('groupChat.tailMessageCount') }}</label>
+                        <NInputNumber v-model:value="compressionConfig.tailMessageCount" :min="1" :step="5" style="width: 100%" />
+                        <p class="form-hint">{{ t('groupChat.tailMessageCountDesc') }}</p>
+                    </div>
+                    <div style="margin-top: 8px">
+                        <NButton
+                            block
+                            :disabled="isCompressing || store.contextStatuses.size > 0"
+                            :loading="isCompressing"
+                            @click="handleForceCompress"
+                        >
+                            {{ isCompressing ? t('groupChat.compressingInProgress') : t('groupChat.compressNow') }}
+                        </NButton>
+                    </div>
+                    <div class="modal-actions">
+                        <NSpace justify="end">
+                            <NButton @click="showCompressionModal = false">{{ t('common.cancel') }}</NButton>
+                            <NButton type="primary" @click="handleSaveCompressionConfig">{{ t('common.save') }}</NButton>
+                        </NSpace>
+                    </div>
+                </div>
+            </div>
+        </Teleport>
+    </div>
+</template>
+
+<script lang="ts">
+import { defineComponent } from 'vue'
+import CreateRoomForm from './CreateRoomForm.vue'
+
+export default defineComponent({ components: { CreateRoomForm } })
+</script>
+
+<style scoped lang="scss">
+@use "@/styles/variables" as *;
+
+.group-chat-panel {
+    display: flex;
+    height: 100%;
+    overflow: hidden;
+}
+
+// ─── Status Bar ──────────────────────────────────────────
+
+.status-bar {
+    flex-shrink: 0;
+    padding: 6px 20px;
+    background: #ffffff;
+    overflow: hidden;
+}
+
+.context-status-list {
+    display: flex;
+    gap: 8px;
+    overflow-x: auto;
+    flex-wrap: nowrap;
+
+    &::-webkit-scrollbar {
+        height: 0;
+    }
+}
+
+.context-status {
+    display: inline-flex;
+    align-items: center;
+    gap: 6px;
+    padding: 4px 10px;
+    font-size: 12px;
+    color: $text-secondary;
+    background-color: $bg-card-hover;
+    border-radius: $radius-sm;
+
+    .dark & {
+        background-color: rgba(255, 255, 255, 0.06);
+    }
+}
+
+.typing-indicator {
+    display: flex;
+    align-items: center;
+    gap: 6px;
+    font-size: 12px;
+    color: $text-muted;
+}
+
+.typing-dots {
+    display: inline-flex;
+    align-items: center;
+    gap: 2px;
+
+    span {
+        display: block;
+        width: 4px;
+        height: 4px;
+        border-radius: 50%;
+        background-color: $text-muted;
+        animation: typing-bounce 1.2s infinite;
+
+        &:nth-child(2) { animation-delay: 0.2s; }
+        &:nth-child(3) { animation-delay: 0.4s; }
+    }
+}
+
+@keyframes typing-bounce {
+    0%, 60%, 100% { transform: translateY(0); opacity: 0.4; }
+    30% { transform: translateY(-3px); opacity: 1; }
+}
+
+// ─── Room Sidebar ────────────────────────────────────────
+
+.room-sidebar {
+    width: 220px;
+    flex-shrink: 0;
+    background-color: $bg-sidebar;
+    border-right: 1px solid $border-color;
+    display: flex;
+    flex-direction: column;
+}
+
+.sidebar-header {
+    display: flex;
+    align-items: center;
+    justify-content: space-between;
+    padding: 12px;
+    flex-shrink: 0;
+
+    .sidebar-title {
+        font-size: 15px;
+        font-weight: 600;
+        color: $text-primary;
+    }
+
+    .sidebar-actions {
+        display: flex;
+        gap: 4px;
+    }
+}
+
+.room-list {
+    flex: 1;
+    overflow-y: auto;
+    padding: 8px;
+}
+
+.room-item {
+    display: flex;
+    align-items: center;
+    gap: 10px;
+    padding: 10px;
+    border-radius: $radius-sm;
+    cursor: pointer;
+    transition: background-color $transition-fast;
+
+    &:hover {
+        background-color: rgba(var(--accent-primary-rgb), 0.06);
+    }
+
+    &.active {
+        background-color: rgba(var(--accent-primary-rgb), 0.12);
+    }
+
+    .room-icon {
+        color: $text-muted;
+        flex-shrink: 0;
+    }
+
+    .room-info {
+        display: flex;
+        flex-direction: column;
+        min-width: 0;
+        flex: 1;
+    }
+
+    .room-name {
+        font-size: 13px;
+        color: $text-primary;
+        white-space: nowrap;
+        overflow: hidden;
+        text-overflow: ellipsis;
+    }
+
+    .room-code {
+        font-size: 11px;
+        color: $text-muted;
+        font-family: $font-code;
+    }
+
+    .room-tokens {
+        font-size: 11px;
+        color: $text-muted;
+    }
+
+    .room-delete-btn {
+        flex-shrink: 0;
+        display: flex;
+        align-items: center;
+        justify-content: center;
+        width: 24px;
+        height: 24px;
+        border: none;
+        background: transparent;
+        color: $text-muted;
+        cursor: pointer;
+        border-radius: $radius-sm;
+        opacity: 0;
+        transition: opacity $transition-fast, color $transition-fast, background-color $transition-fast;
+
+        &:hover {
+            color: $error;
+            background-color: rgba(var(--error-rgb), 0.1);
+        }
+    }
+
+    &:hover .room-delete-btn {
+        opacity: 1;
+    }
+}
+
+.empty-rooms {
+    padding: 20px 12px;
+    text-align: center;
+    font-size: 13px;
+    color: $text-muted;
+}
+
+// ─── Chat Main ──────────────────────────────────────────
+
+.chat-main {
+    flex: 1;
+    display: flex;
+    flex-direction: column;
+    min-width: 0;
+}
+
+.chat-header {
+    display: flex;
+    align-items: center;
+    gap: 12px;
+    padding: 21px 20px;
+    border-bottom: 1px solid $border-color;
+
+    .room-title-text {
+        font-size: 16px;
+        font-weight: 600;
+        color: $text-primary;
+        flex: 1;
+        white-space: nowrap;
+        overflow: hidden;
+        text-overflow: ellipsis;
+    }
+
+    .header-info {
+        display: flex;
+        align-items: center;
+        gap: 8px;
+        flex-shrink: 0;
+    }
+
+    .member-count {
+        font-size: 12px;
+        color: $text-muted;
+    }
+}
+
+// ─── Header Avatar Stack ──────────────────────────────
+
+.avatar-stack {
+    cursor: pointer;
+}
+
+.avatar-stack-inner {
+    display: flex;
+    align-items: center;
+}
+
+.avatar-stack-item {
+    width: 28px;
+    height: 28px;
+    border-radius: 50%;
+    border: 2px solid $bg-card;
+    margin-left: -12px;
+    overflow: hidden;
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    background-color: $bg-secondary;
+    transition: transform $transition-fast;
+
+    &:first-child {
+        margin-left: 0;
+    }
+
+    &:hover {
+        transform: translateY(-2px);
+        z-index: 100 !important;
+    }
+}
+
+.avatar-stack-more {
+    width: 28px;
+    height: 28px;
+    border-radius: 50%;
+    border: 2px solid $bg-card;
+    margin-left: -12px;
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    background-color: $bg-secondary;
+    font-size: 11px;
+    font-weight: 600;
+    color: $text-secondary;
+}
+
+.agent-avatar {
+    width: 28px;
+    height: 28px;
+    display: flex;
+    align-items: center;
+    justify-content: center;
+
+    :deep(svg) {
+        width: 100%;
+        height: 100%;
+    }
+}
+
+// ─── Agent Popover ─────────────────────────────────────
+
+.agent-popover {
+    max-height: 300px;
+    overflow-y: auto;
+}
+
+.agent-popover-title {
+    font-size: 12px;
+    font-weight: 600;
+    color: $text-muted;
+    padding: 0 0 8px;
+    border-bottom: 1px solid $border-color;
+    margin-bottom: 8px;
+}
+
+.agent-popover-item {
+    display: flex;
+    align-items: center;
+    gap: 8px;
+    padding: 6px 4px;
+    border-radius: $radius-sm;
+    transition: background-color $transition-fast;
+
+    &:hover {
+        background-color: rgba(var(--accent-primary-rgb), 0.06);
+    }
+
+    .agent-popover-info {
+        flex: 1;
+        min-width: 0;
+    }
+
+    .agent-popover-name {
+        display: block;
+        font-size: 13px;
+        color: $text-primary;
+        white-space: nowrap;
+        overflow: hidden;
+        text-overflow: ellipsis;
+    }
+
+    .agent-popover-profile {
+        display: block;
+        font-size: 11px;
+        color: $text-muted;
+        white-space: nowrap;
+        overflow: hidden;
+        text-overflow: ellipsis;
+    }
+
+    .agent-popover-remove {
+        display: flex;
+        align-items: center;
+        justify-content: center;
+        width: 24px;
+        height: 24px;
+        border: none;
+        background: none;
+        border-radius: $radius-sm;
+        color: $text-muted;
+        cursor: pointer;
+        flex-shrink: 0;
+        transition: all $transition-fast;
+
+        &:hover {
+            color: $error;
+            background-color: rgba(200, 50, 50, 0.08);
+        }
+    }
+}
+
+// ─── No Room State ────────────────────────────────────────
+
+.no-room {
+    flex: 1;
+    display: flex;
+    flex-direction: column;
+    align-items: center;
+    justify-content: center;
+    gap: 16px;
+    color: $text-muted;
+
+    .no-room-icon {
+        opacity: 0.3;
+    }
+
+    p {
+        font-size: 14px;
+    }
+}
+
+// ─── Shared ──────────────────────────────────────────────
+
+.icon-btn {
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    width: 32px;
+    height: 32px;
+    border: none;
+    background: none;
+    border-radius: $radius-sm;
+    color: $text-secondary;
+    cursor: pointer;
+    transition: all $transition-fast;
+
+    &:hover {
+        background-color: rgba(var(--accent-primary-rgb), 0.08);
+        color: $text-primary;
+    }
+}
+
+.modal-backdrop {
+    position: fixed;
+    inset: 0;
+    background: rgba(0, 0, 0, 0.4);
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    z-index: 1000;
+}
+
+.modal {
+    background: $bg-card;
+    border-radius: $radius-lg;
+    padding: 24px;
+    width: 400px;
+    max-width: 90vw;
+    box-shadow: 0 8px 32px rgba(0, 0, 0, 0.12);
+
+    h3 {
+        font-size: 16px;
+        font-weight: 600;
+        color: $text-primary;
+        margin: 0 0 20px;
+    }
+}
+
+.form-group {
+    margin-bottom: 16px;
+}
+
+.modal-actions {
+    margin-top: 12px;
+    display: flex;
+    justify-content: flex-end;
+    gap: 8px;
+}
+
+.form-hint {
+    font-size: 11px;
+    color: $text-muted;
+    margin: 4px 0 0;
+}
+
+// ─── Connection Dot ──────────────────────────────────────
+
+.connection-dot {
+    width: 8px;
+    height: 8px;
+    border-radius: 50%;
+    flex-shrink: 0;
+
+    &.connected {
+        background-color: $success;
+        box-shadow: 0 0 6px rgba(var(--success-rgb), 0.5);
+    }
+
+    &.disconnected {
+        background-color: $error;
+    }
+}
+
+// ─── Mobile ──────────────────────────────────────────────
+
+@media (max-width: $breakpoint-mobile) {
+    .room-sidebar {
+        position: absolute;
+        left: 0;
+        top: 0;
+        bottom: 0;
+        z-index: 100;
+        box-shadow: 4px 0 16px rgba(0, 0, 0, 0.1);
+    }
+}
+</style>
diff --git a/packages/client/src/components/hermes/group-chat/GroupMessageItem.vue b/packages/client/src/components/hermes/group-chat/GroupMessageItem.vue
new file mode 100644
index 0000000..96027e8
--- /dev/null
+++ b/packages/client/src/components/hermes/group-chat/GroupMessageItem.vue
@@ -0,0 +1,149 @@
+<script setup lang="ts">
+import { computed } from 'vue'
+import multiavatar from '@multiavatar/multiavatar'
+import MarkdownRenderer from '../chat/MarkdownRenderer.vue'
+import type { ChatMessage, RoomAgent } from '@/api/hermes/group-chat'
+
+const props = defineProps<{
+    message: ChatMessage
+    agents: RoomAgent[]
+    currentUserId?: string
+}>()
+
+const isAgent = computed(() => {
+    return props.agents.some(a => a.agentId === props.message.senderId)
+})
+
+const isSelf = computed(() => {
+    return !!props.currentUserId && props.message.senderId === props.currentUserId
+})
+
+const agentInfo = computed(() => {
+    return props.agents.find(a => a.agentId === props.message.senderId)
+})
+
+const timeStr = computed(() => {
+    const d = new Date(props.message.timestamp)
+    return d.toLocaleTimeString([], { hour: '2-digit', minute: '2-digit' })
+})
+
+const avatarSvg = computed(() => {
+    return multiavatar(props.message.senderName || props.message.senderId)
+})
+
+const mentionNames = computed(() => props.agents.map(a => a.name).filter(Boolean))
+</script>
+
+<template>
+    <div class="group-message" :class="{ agent: isAgent, self: isSelf }">
+        <!-- Avatar -->
+        <div class="avatar">
+            <span v-html="avatarSvg" />
+        </div>
+
+        <div class="msg-body">
+            <div class="msg-header">
+                <span class="sender-name">{{ message.senderName }}</span>
+                <span v-if="isAgent && agentInfo?.description" class="agent-desc">{{ agentInfo.description }}</span>
+                <span class="msg-time">{{ timeStr }}</span>
+            </div>
+            <div class="msg-content" :class="{ 'agent-content': isAgent }">
+                <MarkdownRenderer :content="message.content" :mention-names="mentionNames" />
+            </div>
+        </div>
+    </div>
+</template>
+
+<style scoped lang="scss">
+@use "@/styles/variables" as *;
+
+.group-message {
+    display: flex;
+    gap: 10px;
+    padding: 2px 0;
+
+    &.agent,
+    &.self {
+        flex-direction: row-reverse;
+
+        .msg-body {
+            align-items: flex-end;
+        }
+
+        .msg-header {
+            flex-direction: row-reverse;
+        }
+    }
+
+    &.agent .msg-content.agent-content {
+        background-color: rgba(var(--accent-primary-rgb), 0.06);
+    }
+
+    &.self .msg-content {
+        background-color: rgba(var(--accent-primary-rgb), 0.1);
+    }
+}
+
+.avatar {
+    width: 36px;
+    height: 36px;
+    flex-shrink: 0;
+    margin-top: 2px;
+    overflow: hidden;
+    border-radius: 8px;
+
+    :deep(svg) {
+        width: 36px;
+        height: 36px;
+    }
+}
+
+.msg-body {
+    display: flex;
+    flex-direction: column;
+    min-width: 0;
+    max-width: 70%;
+}
+
+.msg-header {
+    display: flex;
+    align-items: center;
+    gap: 8px;
+    padding-bottom: 2px;
+
+    .sender-name {
+        font-size: 13px;
+        font-weight: 600;
+        color: $text-primary;
+    }
+
+    .agent-desc {
+        font-size: 11px;
+        color: $text-muted;
+        font-style: italic;
+    }
+
+    .msg-time {
+        font-size: 11px;
+        color: $text-muted;
+        margin-left: auto;
+    }
+}
+
+.msg-content {
+    padding: 10px 14px;
+    font-size: 14px;
+    line-height: 1.65;
+    color: $text-primary;
+    border-radius: 10px;
+    background-color: $msg-user-bg;
+    word-break: break-word;
+    overflow-wrap: break-word;
+
+    :deep(.mention-highlight) {
+        color: #409eff;
+        font-weight: 600;
+        cursor: default;
+    }
+}
+</style>
diff --git a/packages/client/src/components/hermes/group-chat/GroupMessageList.vue b/packages/client/src/components/hermes/group-chat/GroupMessageList.vue
new file mode 100644
index 0000000..3777031
--- /dev/null
+++ b/packages/client/src/components/hermes/group-chat/GroupMessageList.vue
@@ -0,0 +1,79 @@
+<script setup lang="ts">
+import { ref, watch, nextTick } from 'vue'
+import { useGroupChatStore } from '@/stores/hermes/group-chat'
+import GroupMessageItem from './GroupMessageItem.vue'
+
+const store = useGroupChatStore()
+const listRef = ref<HTMLDivElement>()
+const isNearBottom = ref(true)
+
+function checkNearBottom(): void {
+    if (!listRef.value) return
+    const { scrollTop, scrollHeight, clientHeight } = listRef.value
+    isNearBottom.value = scrollHeight - scrollTop - clientHeight < 200
+}
+
+function scrollToBottom(): void {
+    if (!listRef.value) return
+    listRef.value.scrollTop = listRef.value.scrollHeight
+}
+
+function handleScroll(): void {
+    checkNearBottom()
+}
+
+watch(() => store.messages.length, async () => {
+    await nextTick()
+    if (isNearBottom.value) {
+        scrollToBottom()
+    }
+})
+
+defineExpose({ scrollToBottom })
+</script>
+
+<template>
+    <div ref="listRef" class="message-list" @scroll="handleScroll">
+        <div v-if="store.sortedMessages.length === 0" class="empty-state">
+            <svg width="32" height="32" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1" stroke-linecap="round" stroke-linejoin="round">
+                <path d="M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z" />
+            </svg>
+            <p>No messages yet</p>
+        </div>
+        <GroupMessageItem
+            v-for="msg in store.sortedMessages"
+            :key="msg.id"
+            :message="msg"
+            :agents="store.agents"
+            :current-user-id="store.userId"
+        />
+    </div>
+</template>
+
+<style scoped lang="scss">
+@use "@/styles/variables" as *;
+
+.message-list {
+    flex: 1;
+    overflow-y: auto;
+    padding: 16px 20px;
+    display: flex;
+    flex-direction: column;
+    gap: 12px;
+}
+
+.empty-state {
+    flex: 1;
+    display: flex;
+    flex-direction: column;
+    align-items: center;
+    justify-content: center;
+    gap: 12px;
+    color: $text-muted;
+    opacity: 0.4;
+
+    p {
+        font-size: 14px;
+    }
+}
+</style>
diff --git a/packages/client/src/components/layout/AppSidebar.vue b/packages/client/src/components/layout/AppSidebar.vue
index f62627f..53ec46b 100644
--- a/packages/client/src/components/layout/AppSidebar.vue
+++ b/packages/client/src/components/layout/AppSidebar.vue
@@ -65,21 +65,39 @@ function openChangelog() {
     </div>
 
     <nav class="sidebar-nav">
-      <!-- Chat (standalone) -->
-      <button class="nav-item" :class="{ active: selectedKey === 'hermes.chat' }" @click="handleNav('hermes.chat')">
-        <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
-          <path d="M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z" />
-        </svg>
-        <span>{{ t("sidebar.chat") }}</span>
-      </button>
-
-      <button class="nav-item" @click="openSessionSearch">
-        <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
-          <circle cx="11" cy="11" r="7" />
-          <path d="m20 20-3.5-3.5" />
-        </svg>
-        <span>{{ t("sidebar.search") }}</span>
-      </button>
+      <!-- Conversation -->
+      <div class="nav-group">
+        <div class="nav-group-label" @click="toggleGroup('conversation')">
+          <span>{{ t("sidebar.groupConversation") }}</span>
+          <svg class="nav-group-arrow" :class="{ collapsed: isGroupCollapsed('conversation') }" width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round">
+            <polyline points="6 9 12 15 18 9" />
+          </svg>
+        </div>
+        <div v-show="!isGroupCollapsed('conversation')">
+          <button class="nav-item" :class="{ active: selectedKey === 'hermes.chat' }" @click="handleNav('hermes.chat')">
+            <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
+              <path d="M21 15a2 2 0 0 1-2 2H7l-4 4V5a2 2 0 0 1 2-2h14a2 2 0 0 1 2 2z" />
+            </svg>
+            <span>{{ t("sidebar.chat") }}</span>
+          </button>
+          <button class="nav-item" :class="{ active: selectedKey === 'hermes.groupChat' }" @click="handleNav('hermes.groupChat')">
+            <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
+              <path d="M17 21v-2a4 4 0 0 0-4-4H5a4 4 0 0 0-4 4v2" />
+              <circle cx="9" cy="7" r="4" />
+              <path d="M23 21v-2a4 4 0 0 0-3-3.87" />
+              <path d="M16 3.13a4 4 0 0 1 0 7.75" />
+            </svg>
+            <span>{{ t("sidebar.groupChat") }}<span class="beta-tag">(beta)</span></span>
+          </button>
+          <button class="nav-item" @click="openSessionSearch">
+            <svg width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round">
+              <circle cx="11" cy="11" r="7" />
+              <path d="m20 20-3.5-3.5" />
+            </svg>
+            <span>{{ t("sidebar.search") }}</span>
+          </button>
+        </div>
+      </div>
 
       <!-- Agent -->
       <div class="nav-group">
@@ -430,6 +448,12 @@ function openChangelog() {
     background-color: rgba(var(--accent-primary-rgb), 0.12);
     color: $accent-primary;
   }
+
+  .beta-tag {
+    font-size: 10px;
+    color: $text-muted;
+    margin-left: 2px;
+  }
 }
 
 .sidebar-footer {
diff --git a/packages/client/src/data/changelog.ts b/packages/client/src/data/changelog.ts
index 2cabc64..5ba0b7d 100644
--- a/packages/client/src/data/changelog.ts
+++ b/packages/client/src/data/changelog.ts
@@ -17,7 +17,7 @@ export const changelog: ChangelogEntry[] = [
   },
   {
     version: '0.4.2',
-    date: '2026-03-20',
+    date: '2026-04-22',
     changes: ['changelog.new_0_4_2_1', 'changelog.new_0_4_2_2', 'changelog.new_0_4_2_3', 'changelog.new_0_4_2_4', 'changelog.new_0_4_2_5'],
   },
   {
diff --git a/packages/client/src/i18n/locales/de.ts b/packages/client/src/i18n/locales/de.ts
index 6632356..0fe7043 100644
--- a/packages/client/src/i18n/locales/de.ts
+++ b/packages/client/src/i18n/locales/de.ts
@@ -48,6 +48,7 @@ export default {
     update: 'Aktualisieren',
     create: 'Erstellen',
     saveFailed: 'Speichern fehlgeschlagen',
+    deleteFailed: 'Loschen fehlgeschlagen',
     ok: 'OK',
     copied: 'Kopiert',
     copy: 'Kopieren',
@@ -76,6 +77,8 @@ export default {
     channels: 'Kanale',
     terminal: 'Terminal',
     files: 'Dateien',
+    groupChat: 'Gruppenchat',
+    groupConversation: 'Konversation',
     settings: 'Einstellungen',
     connected: 'Verbunden',
     disconnected: 'Getrennt',
@@ -579,6 +582,55 @@ export default {
     saveFile: 'Speichern',
   },
 
+  // Gruppenchat
+  groupChat: {
+    title: 'Gruppenchat',
+    createRoom: 'Raum erstellen',
+    joinByCode: 'Mit Code beitreten',
+    roomName: 'Raumname',
+    roomNamePlaceholder: 'Raumnamen eingeben',
+    inviteCode: 'Einladungscode',
+    autoGenerate: 'Automatisch generieren',
+    noRooms: 'Noch keine Raume',
+    selectOrCreate: 'Wahlen oder erstellen Sie einen Raum, um zu chatten',
+    agents: 'Agenten',
+    addAgent: 'Agent hinzufugen',
+    selectProfile: 'Wahlen Sie ein Profil',
+    agentAdded: 'Agent hinzugefugt',
+    agentAlreadyInRoom: 'Agent ist bereits in diesem Raum',
+    noAgents: 'Keine Agenten in diesem Raum',
+    members: 'Mitglieder',
+    roomCreated: 'Raum erstellt',
+    roomDeleted: 'Raum gelöscht',
+    deleteRoomConfirm: 'Diesen Raum löschen?',
+    you: 'Du',
+    joined: 'Raum beigetreten',
+    joinFailed: 'Beitreten fehlgeschlagen',
+    inputPlaceholder: 'Nachricht eingeben... (Enter zum Senden)',
+    enterCode: 'Einladungscode eingeben',
+    yourName: 'Dein Name',
+    yourNamePlaceholder: 'Gib deinen Anzeigenamen ein',
+    yourDescription: 'Beschreibung (optional)',
+    yourDescriptionPlaceholder: 'Erzahl anderen wer du bist...',
+    agentName: 'Agent-Name',
+    agentNamePlaceholder: 'Benutzerdefinierter Name (leer = Profilname)',
+    agentDesc: 'Agent-Beschreibung',
+    agentDescPlaceholder: 'Beschreiben Sie, was dieser Agent tut...',
+    agentReplying: 'antwortet...',
+    agentCompressing: 'komprimiert Kontext...',
+    compressionSettings: 'Komprimierungseinstellungen',
+    triggerTokens: 'Trigger-Token',
+    triggerTokensDesc: 'Token-Schwelle für Kontextkomprimierung',
+    maxHistoryTokens: 'Max. Verlaufs-Token',
+    maxHistoryTokensDesc: 'Maximale Token für komprimierten Kontext',
+    tailMessageCount: 'Nachrichten am Ende',
+    tailMessageCountDesc: 'Anzahl der letzten Nachrichten, die unverändert bleiben',
+    compressionConfig: 'Komprimierungskonfig',
+    compressNow: 'Jetzt komprimieren',
+    compressingInProgress: 'Komprimierung läuft, bitte warten',
+    compressionSaved: 'Konfiguration gespeichert',
+  },
+
   // Download
   download: {
     downloading: 'Wird heruntergeladen...',
diff --git a/packages/client/src/i18n/locales/en.ts b/packages/client/src/i18n/locales/en.ts
index c0e5c94..751bbab 100644
--- a/packages/client/src/i18n/locales/en.ts
+++ b/packages/client/src/i18n/locales/en.ts
@@ -48,6 +48,7 @@ export default {
     update: 'Update',
     create: 'Create',
     saveFailed: 'Save failed',
+    deleteFailed: 'Delete failed',
     ok: 'OK',
     copied: 'Copied',
     copy: 'Copy',
@@ -79,6 +80,7 @@ export default {
     channels: 'Channels',
     gateways: 'Gateways',
     terminal: 'Terminal',
+    groupChat: 'Group Chat',
     files: 'Files',
     groupConversation: 'Conversation',
     groupPlatform: 'Platform',
@@ -535,6 +537,55 @@ export default {
     processExited: 'Process exited with code {code}',
   },
 
+  // Group Chat
+  groupChat: {
+    title: 'Group Chat',
+    createRoom: 'Create Room',
+    joinByCode: 'Join by Code',
+    roomName: 'Room Name',
+    roomNamePlaceholder: 'Enter room name',
+    inviteCode: 'Invite Code',
+    autoGenerate: 'Auto-generate',
+    noRooms: 'No rooms yet',
+    selectOrCreate: 'Select or create a room to start chatting',
+    agents: 'Agents',
+    addAgent: 'Add Agent',
+    selectProfile: 'Select a profile',
+    agentAdded: 'Agent added',
+    agentAlreadyInRoom: 'Agent already in this room',
+    noAgents: 'No agents in this room',
+    members: 'members',
+    roomCreated: 'Room created',
+    roomDeleted: 'Room deleted',
+    deleteRoomConfirm: 'Delete this room?',
+    you: 'You',
+    joined: 'Joined room',
+    joinFailed: 'Failed to join room',
+    inputPlaceholder: 'Type a message... (Enter to send)',
+    enterCode: 'Enter invite code',
+    yourName: 'Your Name',
+    yourNamePlaceholder: 'Enter your display name',
+    yourDescription: 'Description (optional)',
+    yourDescriptionPlaceholder: 'Tell others who you are...',
+    agentName: 'Agent Name',
+    agentNamePlaceholder: 'Custom name (leave empty to use profile name)',
+    agentDesc: 'Agent Description',
+    agentDescPlaceholder: 'Describe what this agent does...',
+    agentReplying: 'is replying...',
+    agentCompressing: 'is compressing context...',
+    compressionSettings: 'Compression Settings',
+    triggerTokens: 'Trigger Tokens',
+    triggerTokensDesc: 'Token threshold to trigger context compression',
+    maxHistoryTokens: 'Max History Tokens',
+    maxHistoryTokensDesc: 'Maximum tokens for compressed context sent to LLM',
+    tailMessageCount: 'Tail Message Count',
+    tailMessageCountDesc: 'Number of recent messages to keep verbatim after compression',
+    compressionConfig: 'Compression Config',
+    compressionSaved: 'Compression config saved',
+    compressNow: 'Compress Now',
+    compressingInProgress: 'Compression in progress, please wait',
+  },
+
   // Usage
   usage: {
     title: 'Usage Statistics',
diff --git a/packages/client/src/i18n/locales/es.ts b/packages/client/src/i18n/locales/es.ts
index 30c2bce..9c4a280 100644
--- a/packages/client/src/i18n/locales/es.ts
+++ b/packages/client/src/i18n/locales/es.ts
@@ -48,6 +48,7 @@ export default {
     update: 'Actualizar',
     create: 'Crear',
     saveFailed: 'Error al guardar',
+    deleteFailed: 'Error al eliminar',
     ok: 'OK',
     copied: 'Copiado',
     copy: 'Copiar',
@@ -76,6 +77,8 @@ export default {
     channels: 'Canales',
     terminal: 'Terminal',
     files: 'Archivos',
+    groupChat: 'Chat grupal',
+    groupConversation: 'Conversación',
     settings: 'Configuracion',
     connected: 'Conectado',
     disconnected: 'Desconectado',
@@ -579,6 +582,55 @@ export default {
     saveFile: 'Guardar',
   },
 
+  // Chat grupal
+  groupChat: {
+    title: 'Chat grupal',
+    createRoom: 'Crear sala',
+    joinByCode: 'Unirse con codigo',
+    roomName: 'Nombre de la sala',
+    roomNamePlaceholder: 'Ingrese el nombre de la sala',
+    inviteCode: 'Codigo de invitacion',
+    autoGenerate: 'Generar automaticamente',
+    noRooms: 'Aun no hay salas',
+    selectOrCreate: 'Seleccione o cree una sala para comenzar a chatear',
+    agents: 'Agentes',
+    addAgent: 'Agregar agente',
+    selectProfile: 'Seleccione un perfil',
+    agentAdded: 'Agente agregado',
+    agentAlreadyInRoom: 'El agente ya esta en esta sala',
+    noAgents: 'No hay agentes en esta sala',
+    members: 'Miembros',
+    roomCreated: 'Sala creada',
+    roomDeleted: 'Sala eliminada',
+    deleteRoomConfirm: '¿Eliminar esta sala?',
+    you: 'Tú',
+    joined: 'Se unio a la sala',
+    joinFailed: 'Error al unirse a la sala',
+    inputPlaceholder: 'Escriba un mensaje... (Enter para enviar)',
+    enterCode: 'Ingrese el codigo de invitacion',
+    yourName: 'Tu nombre',
+    yourNamePlaceholder: 'Ingresa tu nombre para mostrar',
+    yourDescription: 'Descripcion (opcional)',
+    yourDescriptionPlaceholder: 'Cuentales a los demas quien eres...',
+    agentName: 'Nombre del agente',
+    agentNamePlaceholder: 'Nombre personalizado (vacío = nombre del perfil)',
+    agentDesc: 'Descripción del agente',
+    agentDescPlaceholder: 'Describe lo que hace este agente...',
+    agentReplying: 'está respondiendo...',
+    agentCompressing: 'está comprimiendo contexto...',
+    compressionSettings: 'Configuración de compresión',
+    triggerTokens: 'Tokens de activación',
+    triggerTokensDesc: 'Umbral de tokens para activar la compresión',
+    maxHistoryTokens: 'Tokens máximos de historial',
+    maxHistoryTokensDesc: 'Máximo de tokens para el contexto comprimido',
+    tailMessageCount: 'Mensajes recientes',
+    tailMessageCountDesc: 'Número de mensajes recientes a conservar sin comprimir',
+    compressionConfig: 'Config. de compresión',
+    compressNow: 'Comprimir ahora',
+    compressingInProgress: 'Compresión en progreso',
+    compressionSaved: 'Configuración guardada',
+  },
+
   // Descarga
   download: {
     downloading: 'Descargando...',
diff --git a/packages/client/src/i18n/locales/fr.ts b/packages/client/src/i18n/locales/fr.ts
index 9cae284..11aab41 100644
--- a/packages/client/src/i18n/locales/fr.ts
+++ b/packages/client/src/i18n/locales/fr.ts
@@ -48,6 +48,7 @@ export default {
     update: 'Mettre a jour',
     create: 'Creer',
     saveFailed: 'Echec de l\'enregistrement',
+    deleteFailed: 'Echec de la suppression',
     ok: 'OK',
     copied: 'Copie',
     copy: 'Copier',
@@ -76,6 +77,8 @@ export default {
     channels: 'Canaux',
     terminal: 'Terminal',
     files: 'Fichiers',
+    groupChat: 'Chat de groupe',
+    groupConversation: 'Conversation',
     settings: 'Parametres',
     connected: 'Connecte',
     disconnected: 'Deconnecte',
@@ -579,6 +582,55 @@ export default {
     saveFile: 'Enregistrer',
   },
 
+  // Chat de groupe
+  groupChat: {
+    title: 'Chat de groupe',
+    createRoom: 'Creer un salon',
+    joinByCode: 'Rejoindre avec un code',
+    roomName: 'Nom du salon',
+    roomNamePlaceholder: 'Entrez le nom du salon',
+    inviteCode: "Code d'invitation",
+    autoGenerate: 'Generer automatiquement',
+    noRooms: 'Aucun salon pour le moment',
+    selectOrCreate: 'Selectionnez ou creez un salon pour commencer a discuter',
+    agents: 'Agents',
+    addAgent: 'Ajouter un agent',
+    selectProfile: 'Selectionnez un profil',
+    agentAdded: 'Agent ajoute',
+    agentAlreadyInRoom: "L'agent est deja dans ce salon",
+    noAgents: 'Aucun agent dans ce salon',
+    members: 'Membres',
+    roomCreated: 'Salon cree',
+    roomDeleted: 'Salon supprime',
+    deleteRoomConfirm: 'Supprimer ce salon ?',
+    you: 'Vous',
+    joined: 'Vous avez rejoint le salon',
+    joinFailed: 'Echec de la connexion au salon',
+    inputPlaceholder: 'Tapez un message... (Entree pour envoyer)',
+    enterCode: "Entrez le code d'invitation",
+    yourName: 'Votre nom',
+    yourNamePlaceholder: "Entrez votre nom d'affichage",
+    yourDescription: 'Description (facultatif)',
+    yourDescriptionPlaceholder: 'Decrivez-vous aux autres...',
+    agentName: 'Nom de l\'agent',
+    agentNamePlaceholder: 'Nom personnalise (vide = nom du profil)',
+    agentDesc: 'Description de l\'agent',
+    agentDescPlaceholder: 'Decrivez le role de cet agent...',
+    agentReplying: 'est en train de répondre...',
+    agentCompressing: 'compresse le contexte...',
+    compressionSettings: 'Paramètres de compression',
+    triggerTokens: 'Seuil de tokens',
+    triggerTokensDesc: 'Seuil de tokens pour déclencher la compression',
+    maxHistoryTokens: 'Max tokens historique',
+    maxHistoryTokensDesc: 'Maximum de tokens pour le contexte compressé',
+    tailMessageCount: 'Messages récents',
+    tailMessageCountDesc: 'Nombre de messages récents à conserver',
+    compressionConfig: 'Config. compression',
+    compressNow: 'Comprimer maintenant',
+    compressingInProgress: 'Compression en cours',
+    compressionSaved: 'Configuration enregistrée',
+  },
+
   // Telechargement
   download: {
     downloading: 'Telechargement...',
diff --git a/packages/client/src/i18n/locales/ja.ts b/packages/client/src/i18n/locales/ja.ts
index aa0dbd8..70bec92 100644
--- a/packages/client/src/i18n/locales/ja.ts
+++ b/packages/client/src/i18n/locales/ja.ts
@@ -48,6 +48,7 @@ export default {
     update: '更新',
     create: '作成',
     saveFailed: '保存に失敗しました',
+    deleteFailed: '削除に失敗しました',
     ok: 'OK',
     copied: 'コピーしました',
     copy: 'コピー',
@@ -76,6 +77,8 @@ export default {
     channels: 'チャンネル',
     terminal: 'ターミナル',
     files: 'ファイル',
+    groupChat: 'グループチャット',
+    groupConversation: '会話',
     settings: '設定',
     connected: '接続済み',
     disconnected: '未接続',
@@ -579,6 +582,55 @@ export default {
     saveFile: '保存',
   },
 
+  // グループチャット
+  groupChat: {
+    title: 'グループチャット',
+    createRoom: 'ルームを作成',
+    joinByCode: 'コードで参加',
+    roomName: 'ルーム名',
+    roomNamePlaceholder: 'ルーム名を入力',
+    inviteCode: '招待コード',
+    autoGenerate: '自動生成',
+    noRooms: 'ルームがありません',
+    selectOrCreate: 'ルームを選択または作成してチャットを開始',
+    agents: 'エージェント',
+    addAgent: 'エージェントを追加',
+    selectProfile: 'プロファイルを選択',
+    agentAdded: 'エージェントが追加されました',
+    agentAlreadyInRoom: 'このエージェントは既にルームにいます',
+    noAgents: 'このルームにエージェントはいません',
+    members: 'メンバー',
+    roomCreated: 'ルームが作成されました',
+    roomDeleted: 'ルームを削除しました',
+    deleteRoomConfirm: 'このルームを削除しますか？',
+    you: 'あなた',
+    joined: 'ルームに参加しました',
+    joinFailed: 'ルームへの参加に失敗しました',
+    inputPlaceholder: 'メッセージを入力... (Enterで送信)',
+    enterCode: '招待コードを入力',
+    yourName: 'あなたの名前',
+    yourNamePlaceholder: '表示名を入力',
+    yourDescription: '自己紹介（任意）',
+    yourDescriptionPlaceholder: '自分について教えてください...',
+    agentName: 'エージェント名',
+    agentNamePlaceholder: 'カスタム名（空欄ならプロファイル名）',
+    agentDesc: 'エージェントの説明',
+    agentDescPlaceholder: 'このエージェントの役割を説明...',
+    agentReplying: 'が返信中...',
+    agentCompressing: 'がコンテキストを圧縮中...',
+    compressionSettings: '圧縮設定',
+    triggerTokens: '圧縮トリガートークン数',
+    triggerTokensDesc: 'このトークン数を超えるとコンテキスト圧縮がトリガーされます',
+    maxHistoryTokens: '最大履歴トークン数',
+    maxHistoryTokensDesc: '圧縮後のLLM送信最大トークン数',
+    tailMessageCount: '末尾メッセージ数',
+    tailMessageCountDesc: '圧縮後にそのまま保持する最近のメッセージ数',
+    compressionConfig: '圧縮設定',
+    compressNow: '今すぐ圧縮',
+    compressingInProgress: '圧縮中です、お待ちください',
+    compressionSaved: '圧縮設定を保存しました',
+  },
+
   // ダウンロード
   download: {
     downloading: 'ダウンロード中...',
diff --git a/packages/client/src/i18n/locales/ko.ts b/packages/client/src/i18n/locales/ko.ts
index 7c5038f..e3fbacf 100644
--- a/packages/client/src/i18n/locales/ko.ts
+++ b/packages/client/src/i18n/locales/ko.ts
@@ -48,6 +48,7 @@ export default {
     update: '업데이트',
     create: '생성',
     saveFailed: '저장 실패',
+    deleteFailed: '삭제 실패',
     ok: '확인',
     copied: '복사됨',
     copy: '복사',
@@ -76,6 +77,8 @@ export default {
     channels: '채널',
     terminal: '터미널',
     files: '파일',
+    groupChat: '그룹 채팅',
+    groupConversation: '대화',
     settings: '설정',
     connected: '연결됨',
     disconnected: '연결 끊김',
@@ -579,6 +582,55 @@ export default {
     saveFile: '저장',
   },
 
+  // 그룹 채팅
+  groupChat: {
+    title: '그룹 채팅',
+    createRoom: '방 만들기',
+    joinByCode: '코드로 참여',
+    roomName: '방 이름',
+    roomNamePlaceholder: '방 이름을 입력하세요',
+    inviteCode: '초대 코드',
+    autoGenerate: '자동 생성',
+    noRooms: '아직 방이 없습니다',
+    selectOrCreate: '방을 선택하거나 만들어 채팅을 시작하세요',
+    agents: '에이전트',
+    addAgent: '에이전트 추가',
+    selectProfile: '프로필 선택',
+    agentAdded: '에이전트가 추가되었습니다',
+    agentAlreadyInRoom: '해당 에이전트가 이미 방에 있습니다',
+    noAgents: '이 방에 에이전트가 없습니다',
+    members: '멤버',
+    roomCreated: '방이 생성되었습니다',
+    roomDeleted: '방이 삭제되었습니다',
+    deleteRoomConfirm: '이 방을 삭제하시겠습니까?',
+    you: '나',
+    joined: '방에 참여했습니다',
+    joinFailed: '방 참여에 실패했습니다',
+    inputPlaceholder: '메시지를 입력하세요... (Enter로 전송)',
+    enterCode: '초대 코드를 입력하세요',
+    yourName: '이름',
+    yourNamePlaceholder: '표시 이름을 입력하세요',
+    yourDescription: '설명 (선택)',
+    yourDescriptionPlaceholder: '자신을 소개해 주세요...',
+    agentName: '에이전트 이름',
+    agentNamePlaceholder: '사용자 지정 이름 (빈칸=프로필 이름)',
+    agentDesc: '에이전트 설명',
+    agentDescPlaceholder: '이 에이전트가 하는 일을 설명...',
+    agentReplying: '이(가) 응답 중...',
+    agentCompressing: '이(가) 컨텍스트 압축 중...',
+    compressionSettings: '압축 설정',
+    triggerTokens: '압축 트리거 토큰',
+    triggerTokensDesc: '이 토큰 수를 초과하면 컨텍스트 압축이 시작됩니다',
+    maxHistoryTokens: '최대 기록 토큰',
+    maxHistoryTokensDesc: '압축된 컨텍스트의 최대 토큰 수',
+    tailMessageCount: '최근 메시지 수',
+    tailMessageCountDesc: '압축 후 그대로 유지할 최근 메시지 수',
+    compressionConfig: '압축 설정',
+    compressNow: '지금 압축',
+    compressingInProgress: '압축 진행 중',
+    compressionSaved: '압축 설정이 저장되었습니다',
+  },
+
   // 다운로드
   download: {
     downloading: '다운로드 중...',
diff --git a/packages/client/src/i18n/locales/pt.ts b/packages/client/src/i18n/locales/pt.ts
index b36d398..9d27915 100644
--- a/packages/client/src/i18n/locales/pt.ts
+++ b/packages/client/src/i18n/locales/pt.ts
@@ -48,6 +48,7 @@ export default {
     update: 'Atualizar',
     create: 'Criar',
     saveFailed: 'Falha ao salvar',
+    deleteFailed: 'Falha ao excluir',
     ok: 'OK',
     copied: 'Copiado',
     copy: 'Copiar',
@@ -76,6 +77,8 @@ export default {
     channels: 'Canais',
     terminal: 'Terminal',
     files: 'Arquivos',
+    groupChat: 'Chat em grupo',
+    groupConversation: 'Conversa',
     settings: 'Configuracoes',
     connected: 'Conectado',
     disconnected: 'Desconectado',
@@ -579,6 +582,55 @@ export default {
     saveFile: 'Salvar',
   },
 
+  // Chat em grupo
+  groupChat: {
+    title: 'Chat em grupo',
+    createRoom: 'Criar sala',
+    joinByCode: 'Entrar com codigo',
+    roomName: 'Nome da sala',
+    roomNamePlaceholder: 'Digite o nome da sala',
+    inviteCode: 'Codigo de convite',
+    autoGenerate: 'Gerar automaticamente',
+    noRooms: 'Nenhuma sala ainda',
+    selectOrCreate: 'Selecione ou crie uma sala para comecar a conversar',
+    agents: 'Agentes',
+    addAgent: 'Adicionar agente',
+    selectProfile: 'Selecione um perfil',
+    agentAdded: 'Agente adicionado',
+    agentAlreadyInRoom: 'O agente ja esta nesta sala',
+    noAgents: 'Nenhum agente nesta sala',
+    members: 'Membros',
+    roomCreated: 'Sala criada',
+    roomDeleted: 'Sala excluída',
+    deleteRoomConfirm: 'Excluir esta sala?',
+    you: 'Você',
+    joined: 'Entrou na sala',
+    joinFailed: 'Falha ao entrar na sala',
+    inputPlaceholder: 'Digite uma mensagem... (Enter para enviar)',
+    enterCode: 'Digite o codigo de convite',
+    yourName: 'Seu nome',
+    yourNamePlaceholder: 'Digite seu nome de exibicao',
+    yourDescription: 'Descricao (opcional)',
+    yourDescriptionPlaceholder: 'Conte aos outros quem voce e...',
+    agentName: 'Nome do agente',
+    agentNamePlaceholder: 'Nome personalizado (vazio = nome do perfil)',
+    agentDesc: 'Descrição do agente',
+    agentDescPlaceholder: 'Descreva o que este agente faz...',
+    agentReplying: 'está respondendo...',
+    agentCompressing: 'está compactando contexto...',
+    compressionSettings: 'Configuração de compactação',
+    triggerTokens: 'Tokens de acionamento',
+    triggerTokensDesc: 'Limite de tokens para acionar a compactação',
+    maxHistoryTokens: 'Máx. tokens de histórico',
+    maxHistoryTokensDesc: 'Máximo de tokens para o contexto compactado',
+    tailMessageCount: 'Mensagens recentes',
+    tailMessageCountDesc: 'Número de mensagens recentes a manter',
+    compressionConfig: 'Config. de compactação',
+    compressNow: 'Compactar agora',
+    compressingInProgress: 'Compactação em andamento',
+    compressionSaved: 'Configuração salva',
+  },
+
   // Download
   download: {
     downloading: 'Baixando...',
diff --git a/packages/client/src/i18n/locales/zh.ts b/packages/client/src/i18n/locales/zh.ts
index e5125f9..cb8aa1a 100644
--- a/packages/client/src/i18n/locales/zh.ts
+++ b/packages/client/src/i18n/locales/zh.ts
@@ -46,6 +46,7 @@ export default {
     save: '保存',
     saved: '已保存',
     saveFailed: '保存失败',
+    deleteFailed: '删除失败',
     ok: '确定',
     copied: '已复制',
     copy: '复制',
@@ -79,6 +80,7 @@ export default {
     channels: '频道',
     gateways: '网关',
     terminal: '终端',
+    groupChat: '群聊',
     files: '文件',
     groupConversation: '对话',
     groupPlatform: '平台',
@@ -537,6 +539,55 @@ export default {
     processExited: '进程已退出，代码 {code}',
   },
 
+  // 群聊
+  groupChat: {
+    title: '群聊',
+    createRoom: '创建房间',
+    joinByCode: '通过邀请码加入',
+    roomName: '房间名称',
+    roomNamePlaceholder: '输入房间名称',
+    inviteCode: '邀请码',
+    autoGenerate: '自动生成',
+    noRooms: '暂无房间',
+    selectOrCreate: '选择或创建一个房间开始聊天',
+    agents: '智能体',
+    addAgent: '添加智能体',
+    selectProfile: '选择一个配置',
+    agentAdded: '智能体已添加',
+    agentAlreadyInRoom: '该智能体已在房间中',
+    noAgents: '当前房间暂无智能体',
+    members: '成员',
+    roomCreated: '房间已创建',
+    roomDeleted: '房间已删除',
+    deleteRoomConfirm: '确定删除这个房间吗？',
+    you: '你',
+    joined: '已加入房间',
+    joinFailed: '加入房间失败',
+    inputPlaceholder: '输入消息... (Enter 发送)',
+    enterCode: '输入邀请码',
+    yourName: '你的名称',
+    yourNamePlaceholder: '输入你的群聊昵称',
+    yourDescription: '自我描述（选填）',
+    yourDescriptionPlaceholder: '介绍一下你自己...',
+    agentName: 'Agent 名称',
+    agentNamePlaceholder: '自定义名称（留空则使用 profile 名称）',
+    agentDesc: 'Agent 描述',
+    agentDescPlaceholder: '描述这个 agent 的作用...',
+    agentReplying: '正在回复...',
+    agentCompressing: '正在压缩上下文...',
+    compressionSettings: '压缩设置',
+    triggerTokens: '触发压缩 Token 数',
+    triggerTokensDesc: '消息 token 数超过此值时触发上下文压缩',
+    maxHistoryTokens: '最大历史 Token 数',
+    maxHistoryTokensDesc: '压缩后发送给 LLM 的最大 token 数',
+    tailMessageCount: '保留最近消息数',
+    tailMessageCountDesc: '压缩后保留最近的原始消息条数',
+    compressionConfig: '压缩配置',
+    compressionSaved: '压缩配置已保存',
+    compressNow: '立即压缩',
+    compressingInProgress: '正在压缩中，请稍后',
+  },
+
   // 用量统计
   usage: {
     title: '用量统计',
diff --git a/packages/client/src/router/index.ts b/packages/client/src/router/index.ts
index bd8b7c5..f2bc1db 100644
--- a/packages/client/src/router/index.ts
+++ b/packages/client/src/router/index.ts
@@ -70,6 +70,11 @@ const router = createRouter({
       name: 'hermes.terminal',
       component: () => import('@/views/hermes/TerminalView.vue'),
     },
+    {
+      path: '/hermes/group-chat',
+      name: 'hermes.groupChat',
+      component: () => import('@/views/hermes/GroupChatView.vue'),
+    },
     {
       path: '/hermes/files',
       name: 'hermes.files',
diff --git a/packages/client/src/stores/hermes/chat.ts b/packages/client/src/stores/hermes/chat.ts
index d0210cd..8355dbf 100644
--- a/packages/client/src/stores/hermes/chat.ts
+++ b/packages/client/src/stores/hermes/chat.ts
@@ -492,7 +492,16 @@ export const useChatStore = defineStore('chat', () => {
       // that was just created and whose first run is still in-flight. Without
       // this, refreshing mid-run would wipe the session and fall back to
       // sessions[0], which is exactly what the user reported.
-      const localOnly = sessions.value.filter(s => !freshIds.has(s.id))
+      // Sessions without an active in-flight run are considered deleted and
+      // cleaned up along with their cached messages.
+      const localOnly = sessions.value.filter(s => {
+        if (freshIds.has(s.id)) return false
+        if (readInFlight(s.id)) return true
+        // Session no longer exists on server and no active run — clean up cache
+        removeItemWithLegacy(msgsCacheKey(s.id), legacyMsgsCacheKey(s.id))
+        removeItemWithLegacy(inFlightKey(s.id), legacyInFlightKey(s.id))
+        return false
+      })
       sessions.value = [...localOnly, ...fresh]
       persistSessionsList()
 
diff --git a/packages/client/src/stores/hermes/group-chat.ts b/packages/client/src/stores/hermes/group-chat.ts
new file mode 100644
index 0000000..9b907f9
--- /dev/null
+++ b/packages/client/src/stores/hermes/group-chat.ts
@@ -0,0 +1,365 @@
+import { defineStore } from 'pinia'
+import { ref, computed } from 'vue'
+import {
+    connectGroupChat,
+    disconnectGroupChat,
+    getSocket,
+    getStoredUserId,
+    getStoredUserName,
+    type RoomInfo,
+    type RoomAgent,
+    type ChatMessage,
+    type MemberInfo,
+    createRoom,
+    listRooms,
+    getRoomDetail,
+    joinRoomByCode,
+    addAgent,
+    listAgents,
+    removeAgent,
+    deleteRoom as deleteRoomApi,
+} from '@/api/hermes/group-chat'
+
+export const useGroupChatStore = defineStore('groupChat', () => {
+    // ─── State ─────────────────────────────────────────────
+    const connected = ref(false)
+    const currentRoomId = ref<string | null>(null)
+    const rooms = ref<RoomInfo[]>([])
+    const messages = ref<ChatMessage[]>([])
+    const members = ref<MemberInfo[]>([])
+    const agents = ref<RoomAgent[]>([])
+    const roomName = ref('')
+    const isJoining = ref(false)
+    const error = ref<string | null>(null)
+    const typingUsers = ref<Map<string, { name: string; timer: ReturnType<typeof setTimeout> }>>(new Map())
+    const contextStatuses = ref<Map<string, { agentName: string; status: string }>>(new Map())
+
+    // Computed: returns first active status for backward compat
+    const contextStatus = computed(() => {
+        for (const [, status] of contextStatuses.value) {
+            return status
+        }
+        return null
+    })
+    const userId = ref(getStoredUserId())
+    const userName = ref(getStoredUserName() || '')
+
+    // ─── Computed ───────────────────────────────────────────
+    const sortedMessages = computed(() => {
+        return [...messages.value].sort((a, b) => a.timestamp - b.timestamp)
+    })
+
+    const memberNames = computed(() => {
+        return members.value.map(m => m.name)
+    })
+
+    const typingNames = computed(() => {
+        return Array.from(typingUsers.value.values()).map(u => u.name)
+    })
+
+    const typingText = computed(() => {
+        const names = typingNames.value
+        if (names.length === 0) return ''
+        if (names.length === 1) return `${names[0]} is typing...`
+        if (names.length === 2) return `${names[0]} and ${names[1]} are typing...`
+        return `${names[0]} and ${names.length - 1} others are typing...`
+    })
+
+    // ─── Connection ────────────────────────────────────────
+    function connect() {
+        const socket = connectGroupChat({
+            userId: userId.value,
+            userName: userName.value || undefined,
+        })
+        console.log('[GroupChat] connecting...', { userId: userId.value, userName: userName.value })
+
+        socket.on('connect', () => {
+            console.log('[GroupChat] connected, socket id:', socket.id)
+            connected.value = true
+            error.value = null
+        })
+
+        socket.on('disconnect', (reason) => {
+            console.log('[GroupChat] disconnected:', reason)
+            connected.value = false
+        })
+
+        socket.on('connect_error', (err: Error) => {
+            console.error('[GroupChat] connect_error:', err.message)
+            error.value = err.message
+            connected.value = false
+        })
+
+        socket.on('message', (msg: ChatMessage) => {
+            if (msg.roomId === currentRoomId.value) {
+                messages.value.push(msg)
+            }
+        })
+
+        socket.on('member_joined', (data: { roomId: string; members: MemberInfo[] }) => {
+            if (data.roomId === currentRoomId.value) {
+                members.value = data.members
+            }
+        })
+
+        socket.on('member_left', (data: { roomId: string; members: MemberInfo[] }) => {
+            if (data.roomId === currentRoomId.value) {
+                members.value = data.members
+            }
+        })
+
+        socket.on('typing', (data: { roomId: string; userId: string; userName: string }) => {
+            if (data.roomId === currentRoomId.value && !typingUsers.value.has(data.userId)) {
+                const timer = setTimeout(() => typingUsers.value.delete(data.userId), 5000)
+                typingUsers.value.set(data.userId, { name: data.userName, timer })
+            }
+        })
+
+        socket.on('stop_typing', (data: { roomId: string; userId: string }) => {
+            if (data.roomId === currentRoomId.value && typingUsers.value.has(data.userId)) {
+                const entry = typingUsers.value.get(data.userId)!
+                clearTimeout(entry.timer)
+                typingUsers.value.delete(data.userId)
+            }
+        })
+
+        socket.on('context_status', (data: { roomId: string; agentName: string; status: string }) => {
+            if (data.roomId === currentRoomId.value) {
+                if (data.status === 'ready') {
+                    contextStatuses.value.delete(data.agentName)
+                } else {
+                    contextStatuses.value.set(data.agentName, { agentName: data.agentName, status: data.status })
+                }
+                // Trigger reactivity
+                contextStatuses.value = new Map(contextStatuses.value)
+            }
+        })
+
+        socket.on('room_updated', (data: { roomId: string; totalTokens: number }) => {
+            const room = rooms.value.find(r => r.id === data.roomId)
+            if (room) room.totalTokens = data.totalTokens
+        })
+    }
+
+    function disconnect() {
+        disconnectGroupChat()
+        connected.value = false
+        currentRoomId.value = null
+        messages.value = []
+        members.value = []
+        agents.value = []
+        roomName.value = ''
+        typingUsers.value.clear()
+        contextStatuses.value.clear()
+    }
+
+    function setUserInfo(name: string, description: string) {
+        userName.value = name
+        localStorage.setItem('gc_user_name', name)
+        localStorage.setItem('gc_user_description', description)
+    }
+
+    // ─── Room Actions ──────────────────────────────────────
+    async function joinRoom(roomId: string) {
+        isJoining.value = true
+        error.value = null
+
+        try {
+            const res = await getRoomDetail(roomId)
+            currentRoomId.value = res.room.id
+            roomName.value = res.room.name
+            messages.value = res.messages
+            agents.value = res.agents
+            members.value = res.members || []
+        } catch (err: any) {
+            error.value = err.message
+            throw err
+        } finally {
+            isJoining.value = false
+        }
+
+        // Join via socket for real-time updates
+        const socket = getSocket()
+        if (socket) {
+            await new Promise<void>((resolve) => {
+                socket.emit('join', { roomId, name: userName.value || undefined }, (res: any) => {
+                    if (!res?.error) {
+                        members.value = res.members || []
+                        if (res.agents) agents.value = res.agents
+
+                        // Restore typing state from server
+                        if (res.typingUsers) {
+                            for (const u of res.typingUsers) {
+                                if (!typingUsers.value.has(u.userId)) {
+                                    const timer = setTimeout(() => typingUsers.value.delete(u.userId), 5000)
+                                    typingUsers.value.set(u.userId, { name: u.userName, timer })
+                                }
+                            }
+                        }
+
+                        // Restore context statuses from server
+                        if (res.contextStatuses) {
+                            contextStatuses.value = new Map(
+                                res.contextStatuses.map((s: any) => [s.agentName, s])
+                            )
+                        }
+                    }
+                    resolve()
+                })
+            })
+        }
+    }
+
+    async function sendMessage(content: string) {
+        const socket = getSocket()
+        if (!socket || !currentRoomId.value) return
+        emitStopTyping()
+
+        return new Promise<void>((resolve, reject) => {
+            socket!.emit('message', { roomId: currentRoomId.value, content }, (res: { id?: string; error?: string }) => {
+                if (res.error) {
+                    reject(new Error(res.error))
+                    return
+                }
+                resolve()
+            })
+        })
+    }
+
+    async function loadRooms() {
+        try {
+            const res = await listRooms()
+            rooms.value = res.rooms
+        } catch (err: any) {
+            error.value = err.message
+        }
+    }
+
+    async function createNewRoom(name: string, inviteCode: string, agentList?: { profile: string; name?: string; description?: string; invited?: boolean }[], compression?: { triggerTokens: number; maxHistoryTokens: number; tailMessageCount: number }) {
+        try {
+            const res = await createRoom({
+                name,
+                inviteCode,
+                agents: agentList,
+                compression: compression || { triggerTokens: 100000, maxHistoryTokens: 32000, tailMessageCount: 20 },
+            })
+            rooms.value.push(res.room)
+            return res
+        } catch (err: any) {
+            error.value = err.message
+            throw err
+        }
+    }
+
+    async function joinByCode(code: string) {
+        try {
+            const res = await joinRoomByCode(code)
+            await joinRoom(res.room.id)
+            return res.room
+        } catch (err: any) {
+            error.value = err.message
+            throw err
+        }
+    }
+
+    async function deleteRoom(roomId: string) {
+        try {
+            await deleteRoomApi(roomId)
+            rooms.value = rooms.value.filter(r => r.id !== roomId)
+            if (currentRoomId.value === roomId) {
+                currentRoomId.value = null
+                messages.value = []
+                members.value = []
+                agents.value = []
+                roomName.value = ''
+            }
+        } catch (err: any) {
+            error.value = err.message
+            throw err
+        }
+    }
+
+    // ─── Agent Actions ─────────────────────────────────────
+    async function loadAgents(roomId: string) {
+        try {
+            const res = await listAgents(roomId)
+            agents.value = res.agents
+        } catch { /* ignore */ }
+    }
+
+    async function addAgentToRoom(roomId: string, data: { profile: string; name?: string; description?: string; invited?: boolean }) {
+        try {
+            const res = await addAgent(roomId, data)
+            agents.value.push(res.agent)
+            return res.agent
+        } catch (err: any) {
+            error.value = err.message
+            throw err
+        }
+    }
+
+    async function removeAgentFromRoom(roomId: string, agentId: string) {
+        try {
+            await removeAgent(roomId, agentId)
+            agents.value = agents.value.filter(a => a.id !== agentId)
+        } catch (err: any) {
+            error.value = err.message
+            throw err
+        }
+    }
+
+    // ─── Typing ────────────────────────────────────────────
+    let _typingTimer: ReturnType<typeof setTimeout> | null = null
+
+    function emitTyping() {
+        const socket = getSocket()
+        if (!socket || !currentRoomId.value) return
+        socket.emit('typing', { roomId: currentRoomId.value })
+        if (_typingTimer) clearTimeout(_typingTimer)
+        _typingTimer = setTimeout(() => emitStopTyping(), 4000)
+    }
+
+    function emitStopTyping() {
+        const socket = getSocket()
+        if (!socket || !currentRoomId.value) return
+        socket.emit('stop_typing', { roomId: currentRoomId.value })
+        if (_typingTimer) { clearTimeout(_typingTimer); _typingTimer = null }
+    }
+
+    return {
+        // State
+        connected,
+        currentRoomId,
+        rooms,
+        messages,
+        members,
+        agents,
+        roomName,
+        isJoining,
+        error,
+        contextStatus,
+        contextStatuses,
+        userId,
+        userName,
+        // Computed
+        sortedMessages,
+        memberNames,
+        typingNames,
+        typingText,
+        // Actions
+        connect,
+        disconnect,
+        setUserInfo,
+        joinRoom,
+        sendMessage,
+        loadRooms,
+        emitTyping,
+        emitStopTyping,
+        createNewRoom,
+        joinByCode,
+        deleteRoom,
+        loadAgents,
+        addAgentToRoom,
+        removeAgentFromRoom,
+    }
+})
diff --git a/packages/client/src/views/hermes/GroupChatView.vue b/packages/client/src/views/hermes/GroupChatView.vue
new file mode 100644
index 0000000..c9b6881
--- /dev/null
+++ b/packages/client/src/views/hermes/GroupChatView.vue
@@ -0,0 +1,30 @@
+<script setup lang="ts">
+import { onMounted, onUnmounted } from 'vue'
+import GroupChatPanel from '@/components/hermes/group-chat/GroupChatPanel.vue'
+import { useGroupChatStore } from '@/stores/hermes/group-chat'
+
+const store = useGroupChatStore()
+
+onMounted(() => {
+    store.connect()
+    store.loadRooms()
+})
+
+onUnmounted(() => {
+    store.disconnect()
+})
+</script>
+
+<template>
+    <div class="group-chat-view">
+        <GroupChatPanel />
+    </div>
+</template>
+
+<style scoped lang="scss">
+.group-chat-view {
+    height: calc(100 * var(--vh));
+    display: flex;
+    flex-direction: column;
+}
+</style>
diff --git a/packages/server/src/controllers/hermes/profiles.ts b/packages/server/src/controllers/hermes/profiles.ts
index feca671..e4dd001 100644
--- a/packages/server/src/controllers/hermes/profiles.ts
+++ b/packages/server/src/controllers/hermes/profiles.ts
@@ -3,6 +3,7 @@ import { mkdir, writeFile } from 'fs/promises'
 import { basename, join } from 'path'
 import { tmpdir } from 'os'
 import * as hermesCli from '../../services/hermes/hermes-cli'
+import { drainPendingSessionDeletes } from '../../services/hermes/group-chat'
 import { getGatewayManagerInstance } from '../../services/gateway-bootstrap'
 import { logger } from '../../services/logger'
 
@@ -118,7 +119,17 @@ export async function switchProfile(ctx: any) {
     } catch (err: any) {
       logger.error(err, 'Ensure config failed')
     }
-    ctx.body = { success: true, message: output.trim() }
+    const drainResult = await drainPendingSessionDeletes(name)
+    logger.info('[switchProfile] drain result for profile "%s": %d deleted, %d failed', name, drainResult.deleted.length, drainResult.failed.length)
+    if (drainResult.failed.length > 0) {
+      logger.warn({ profile: name, failed: drainResult.failed }, 'Failed to drain some pending session deletes after profile switch')
+    }
+    ctx.body = {
+      success: true,
+      message: output.trim(),
+      drained_session_deletes: drainResult.deleted.length,
+      failed_session_deletes: drainResult.failed.length,
+    }
   } catch (err: any) {
     ctx.status = 500
     ctx.body = { error: err.message }
diff --git a/packages/server/src/controllers/hermes/sessions.ts b/packages/server/src/controllers/hermes/sessions.ts
index 1fad54a..ac21dd8 100644
--- a/packages/server/src/controllers/hermes/sessions.ts
+++ b/packages/server/src/controllers/hermes/sessions.ts
@@ -7,6 +7,9 @@ import {
 import { listSessionSummaries, searchSessionSummaries } from '../../db/hermes/sessions-db'
 import { deleteUsage, getUsage, getUsageBatch } from '../../db/hermes/usage-store'
 import { getModelContextLength } from '../../services/hermes/model-context'
+import type { ConversationDetail, ConversationSummary } from '../../services/hermes/conversations'
+import { getActiveProfileName } from '../../services/hermes/hermes-profile'
+import { getGroupChatServer } from '../../routes/hermes/group-chat'
 import { logger } from '../../services/logger'
 
 function parseHumanOnly(value: unknown): boolean {
@@ -20,6 +23,35 @@ function parseLimit(value: unknown): number | undefined {
   return Number.isFinite(parsed) && parsed > 0 ? parsed : undefined
 }
 
+function getPendingDeletedSessionIds(): Set<string> {
+  return getGroupChatServer()?.getStorage().getPendingDeletedSessionIds() || new Set<string>()
+}
+
+function isPendingDeletedSession(sessionId: string): boolean {
+  return getPendingDeletedSessionIds().has(sessionId)
+}
+
+function filterPendingDeletedSessions<T extends { id: string }>(items: T[]): T[] {
+  const pendingIds = getPendingDeletedSessionIds()
+  if (pendingIds.size === 0) return items
+  return items.filter(item => !pendingIds.has(item.id))
+}
+
+function filterPendingDeletedConversationSummaries(items: ConversationSummary[]): ConversationSummary[] {
+  return filterPendingDeletedSessions(items)
+}
+
+function hasPendingDeletedConversation(detail: ConversationDetail): boolean {
+  const pendingIds = getPendingDeletedSessionIds()
+  if (pendingIds.size === 0) return false
+  if (pendingIds.has(detail.session_id)) return true
+  return detail.messages.some(message => pendingIds.has(message.session_id))
+}
+
+function getGroupChatStorage() {
+  return getGroupChatServer()?.getStorage() || null
+}
+
 export async function listConversations(ctx: any) {
   const source = (ctx.query.source as string) || undefined
   const humanOnly = parseHumanOnly(ctx.query.humanOnly)
@@ -27,14 +59,14 @@ export async function listConversations(ctx: any) {
 
   try {
     const sessions = await listConversationSummariesFromDb({ source, humanOnly, limit })
-    ctx.body = { sessions }
+    ctx.body = { sessions: filterPendingDeletedConversationSummaries(sessions) }
     return
   } catch (err) {
     logger.warn(err, 'Hermes Conversation DB: summary query failed, falling back to CLI export')
   }
 
   const sessions = await listConversationSummaries({ source, humanOnly, limit })
-  ctx.body = { sessions }
+  ctx.body = { sessions: filterPendingDeletedConversationSummaries(sessions) }
 }
 
 export async function getConversationMessages(ctx: any) {
@@ -43,7 +75,7 @@ export async function getConversationMessages(ctx: any) {
 
   try {
     const detail = await getConversationDetailFromDb(ctx.params.id, { source, humanOnly })
-    if (!detail) {
+    if (!detail || hasPendingDeletedConversation(detail)) {
       ctx.status = 404
       ctx.body = { error: 'Conversation not found' }
       return
@@ -55,7 +87,7 @@ export async function getConversationMessages(ctx: any) {
   }
 
   const detail = await getConversationDetail(ctx.params.id, { source, humanOnly })
-  if (!detail) {
+  if (!detail || hasPendingDeletedConversation(detail)) {
     ctx.status = 404
     ctx.body = { error: 'Conversation not found' }
     return
@@ -69,14 +101,14 @@ export async function list(ctx: any) {
 
   try {
     const sessions = await listSessionSummaries(source, limit && limit > 0 ? limit : 2000)
-    ctx.body = { sessions }
+    ctx.body = { sessions: filterPendingDeletedSessions(sessions) }
     return
   } catch (err) {
     logger.warn(err, 'Hermes Session DB: summary query failed, falling back to CLI')
   }
 
   const sessions = await hermesCli.listSessions(source, limit)
-  ctx.body = { sessions }
+  ctx.body = { sessions: filterPendingDeletedSessions(sessions) }
 }
 
 export async function search(ctx: any) {
@@ -88,7 +120,7 @@ export async function search(ctx: any) {
 
   try {
     const results = await searchSessionSummaries(q, source, limit && limit > 0 ? limit : 20)
-    ctx.body = { results }
+    ctx.body = { results: filterPendingDeletedSessions(results) }
   } catch (err) {
     logger.error(err, 'Hermes Session DB: search failed')
     ctx.status = 500
@@ -97,6 +129,12 @@ export async function search(ctx: any) {
 }
 
 export async function get(ctx: any) {
+  if (isPendingDeletedSession(ctx.params.id)) {
+    ctx.status = 404
+    ctx.body = { error: 'Session not found' }
+    return
+  }
+
   const session = await hermesCli.getSession(ctx.params.id)
   if (!session) {
     ctx.status = 404
@@ -107,14 +145,44 @@ export async function get(ctx: any) {
 }
 
 export async function remove(ctx: any) {
-  const ok = await hermesCli.deleteSession(ctx.params.id)
-  if (!ok) {
-    ctx.status = 500
-    ctx.body = { error: 'Failed to delete session' }
+  const sessionId = ctx.params.id
+  const storage = getGroupChatStorage()
+  const currentProfile = getActiveProfileName()
+  const mapped = storage?.getSessionProfile(sessionId) || null
+
+  logger.info('[remove] sessionId=%s, currentProfile=%s, mapped=%j', sessionId, currentProfile, mapped)
+
+  if (!mapped) {
+    logger.info('[remove] no mapping found, deleting directly')
+    const ok = await hermesCli.deleteSession(sessionId)
+    if (!ok) {
+      ctx.status = 500
+      ctx.body = { error: 'Failed to delete session' }
+      return
+    }
+    deleteUsage(sessionId)
+    ctx.body = { ok: true }
     return
   }
-  deleteUsage(ctx.params.id)
-  ctx.body = { ok: true }
+
+  if (mapped.profile_name === currentProfile) {
+    logger.info('[remove] same profile, deleting directly')
+    const ok = await hermesCli.deleteSession(sessionId)
+    if (!ok) {
+      ctx.status = 500
+      ctx.body = { error: 'Failed to delete session' }
+      return
+    }
+    storage?.deleteSessionProfile(sessionId)
+    deleteUsage(sessionId)
+    ctx.body = { ok: true }
+    return
+  }
+
+  logger.info('[remove] cross-profile detected, enqueued deferred delete for profile=%s', mapped.profile_name)
+  storage?.enqueuePendingSessionDelete(sessionId, mapped.profile_name)
+  deleteUsage(sessionId)
+  ctx.body = { ok: true, deferred: true }
 }
 
 export async function usageBatch(ctx: any) {
diff --git a/packages/server/src/db/hermes/sessions-db.ts b/packages/server/src/db/hermes/sessions-db.ts
index 203e564..950d973 100644
--- a/packages/server/src/db/hermes/sessions-db.ts
+++ b/packages/server/src/db/hermes/sessions-db.ts
@@ -159,20 +159,73 @@ function containsCjk(text: string): boolean {
   return false
 }
 
-function isNumericQuery(text: string): boolean {
-  return /^\d+(?:\s+\d+)*$/.test(text.trim())
+function escapeLikePattern(value: string): string {
+  return value.replace(/[\\%_]/g, (match) => `\\${match}`)
 }
 
-function hasUnsafeChars(text: string): boolean {
-  return /[^\w\s\u4e00-\u9fff\u3400-\u4dbf\u3000-\u303f\u3040-\u309f\u30a0-\u30ff\uac00-\ud7af]/.test(text)
+function buildLikePattern(value: string): string {
+  return `%${escapeLikePattern(value)}%`
 }
 
-function runLikeContentSearch(
+function normalizeTitleLikeQuery(query: string): string {
+  const tokens = query.match(/"[^"]*"\*?|\S+/g)
+  if (!tokens) return query
+
+  const normalizedTokens = tokens
+    .map((token) => {
+      let value = token.endsWith('*') ? token.slice(0, -1) : token
+      if (value.startsWith('"') && value.endsWith('"')) {
+        value = value.slice(1, -1)
+      }
+      return value
+    })
+    .filter(Boolean)
+
+  return normalizedTokens.join(' ').trim() || query
+}
+
+function shouldUseLiteralContentSearch(query: string): boolean {
+  const trimmed = query.trim()
+  if (!trimmed) return false
+  if (/[^\p{L}\p{N}\s"*.-]/u.test(trimmed)) return true
+
+  const tokens = trimmed.match(/"[^"]*"\*?|\S+/g)
+  if (!tokens) return true
+
+  for (const token of tokens) {
+    if (/^(AND|OR|NOT)$/i.test(token)) continue
+
+    const raw = token.endsWith('*') ? token.slice(0, -1) : token
+    if (!raw) return true
+
+    if (raw.startsWith('"') && raw.endsWith('"')) {
+      const inner = raw.slice(1, -1)
+      if (!inner.trim()) return true
+      if (!/^[\p{L}\p{N}\s.-]+$/u.test(inner)) return true
+      if ((inner.includes('.') || inner.includes('-')) && !/^[\p{L}\p{N}]+(?:[.-][\p{L}\p{N}]+)*(?:\s+[\p{L}\p{N}]+(?:[.-][\p{L}\p{N}]+)*)*$/u.test(inner)) return true
+      continue
+    }
+
+    if (raw.includes('.') || raw.includes('-')) {
+      if (!/^[\p{L}\p{N}]+(?:[.-][\p{L}\p{N}]+)*$/u.test(raw)) return true
+      continue
+    }
+
+    if (!/^[\p{L}\p{N}]+$/u.test(raw)) return true
+  }
+
+  return false
+}
+
+function runLiteralContentSearch(
   db: { prepare: (sql: string) => { all: (...params: any[]) => Record<string, unknown>[] } },
   source: string | undefined,
   query: string,
+  limit: number,
 ): Record<string, unknown>[] {
   const likeBase = buildBaseSessionSql(source)
+  const loweredQuery = query.toLowerCase()
+  const likePattern = buildLikePattern(loweredQuery)
   const likeSql = `
     WITH base AS (
       ${likeBase.sql}
@@ -182,17 +235,17 @@ function runLikeContentSearch(
       m.id AS matched_message_id,
       substr(
         m.content,
-        max(1, instr(m.content, ?) - 40),
+        max(1, instr(LOWER(m.content), ?) - 40),
         120
       ) AS snippet,
       0 AS rank
     FROM base
     JOIN messages m ON m.session_id = base.id
-    WHERE m.content LIKE ?
+    WHERE LOWER(m.content) LIKE ? ESCAPE '\\'
     ORDER BY base.last_active DESC, m.timestamp DESC
+    LIMIT ?
   `
-  const likeStatement = db.prepare(likeSql)
-  return likeStatement.all(...likeBase.params, query, `%${query}%`) as Record<string, unknown>[]
+  return db.prepare(likeSql).all(...likeBase.params, loweredQuery, likePattern, limit * 4) as Record<string, unknown>[]
 }
 
 function sanitizeFtsQuery(query: string): string {
@@ -208,7 +261,7 @@ function sanitizeFtsQuery(query: string): string {
   sanitized = sanitized.replace(/(^|\s)\*/g, '$1')
   sanitized = sanitized.trim().replace(/^(AND|OR|NOT)\b\s*/i, '')
   sanitized = sanitized.trim().replace(/\s+(AND|OR|NOT)\s*$/i, '')
-  sanitized = sanitized.replace(/\b(\w+(?:[.-]\w+)+)\b/g, '"$1"')
+  sanitized = sanitized.replace(/\b([\p{L}\p{N}]+(?:[.-][\p{L}\p{N}]+)+)\b/gu, '"$1"')
 
   for (let i = 0; i < quotedParts.length; i += 1) {
     sanitized = sanitized.replace(`\u0000Q${i}\u0000`, quotedParts[i])
@@ -218,7 +271,7 @@ function sanitizeFtsQuery(query: string): string {
 }
 
 function toPrefixQuery(query: string): string {
-  const tokens = query.match(/"[^"]*"|\S+/g)
+  const tokens = query.match(/"[^"]*"\*?|\S+/g)
   if (!tokens) return ''
   return tokens
     .map((token) => {
@@ -282,6 +335,8 @@ export async function searchSessionSummaries(
   const db = new DatabaseSync(sessionDbPath(), { open: true, readOnly: true })
   const normalized = sanitizeFtsQuery(trimmed)
   const prefixQuery = toPrefixQuery(normalized)
+  const titlePattern = buildLikePattern(normalizeTitleLikeQuery(trimmed).toLowerCase())
+  const useLiteralContentSearch = containsCjk(trimmed) || shouldUseLiteralContentSearch(trimmed)
   let titleRows: Record<string, unknown>[] = []
 
   try {
@@ -301,13 +356,13 @@ export async function searchSessionSummaries(
         END AS snippet,
         0 AS rank
       FROM base
-      WHERE LOWER(COALESCE(base.title, '')) LIKE ?
+      WHERE LOWER(COALESCE(base.title, '')) LIKE ? ESCAPE '\\'
       ORDER BY base.last_active DESC
       LIMIT ?
     `
 
     const titleStatement = db.prepare(titleSql)
-    titleRows = titleStatement.all(...titleBase.params, `%${trimmed.toLowerCase()}%`, limit) as Record<string, unknown>[]
+    titleRows = titleStatement.all(...titleBase.params, titlePattern, limit) as Record<string, unknown>[]
 
     const contentSql = `
       WITH base AS (
@@ -326,9 +381,11 @@ export async function searchSessionSummaries(
       LIMIT ?
     `
 
-    const contentRows = prefixQuery
-      ? (db.prepare(contentSql).all(...contentBase.params, prefixQuery, limit * 4) as Record<string, unknown>[])
-      : []
+    const contentRows = useLiteralContentSearch
+      ? runLiteralContentSearch(db, source, trimmed, limit)
+      : prefixQuery
+        ? (db.prepare(contentSql).all(...contentBase.params, prefixQuery, limit * 4) as Record<string, unknown>[])
+        : []
 
     const merged = new Map<string, HermesSessionSearchRow>()
     for (const row of titleRows) {
@@ -351,19 +408,7 @@ export async function searchSessionSummaries(
   } catch (err) {
     const message = err instanceof Error ? err.message : String(err)
     if (containsCjk(normalized)) {
-      const likeRows = runLikeContentSearch(db, source, trimmed)
-      const merged = new Map<string, HermesSessionSearchRow>()
-      for (const row of likeRows) {
-        const mapped = mapSearchRow(row)
-        if (!merged.has(mapped.id)) {
-          merged.set(mapped.id, mapped)
-        }
-      }
-      return [...merged.values()].slice(0, limit)
-    }
-
-    if (isNumericQuery(trimmed) || hasUnsafeChars(trimmed)) {
-      const likeRows = runLikeContentSearch(db, source, trimmed)
+      const likeRows = runLiteralContentSearch(db, source, trimmed, limit)
       const merged = new Map<string, HermesSessionSearchRow>()
       for (const row of titleRows) {
         const mapped = mapSearchRow(row)
@@ -375,7 +420,12 @@ export async function searchSessionSummaries(
           merged.set(mapped.id, mapped)
         }
       }
-      return [...merged.values()].slice(0, limit)
+      const items = [...merged.values()]
+      items.sort((a, b) => {
+        if (a.rank !== b.rank) return a.rank - b.rank
+        return b.last_active - a.last_active
+      })
+      return items.slice(0, limit)
     }
 
     throw new Error(`Failed to search sessions: ${message}`)
diff --git a/packages/server/src/index.ts b/packages/server/src/index.ts
index e9b12a2..c40eae9 100644
--- a/packages/server/src/index.ts
+++ b/packages/server/src/index.ts
@@ -9,11 +9,13 @@ import { mkdir } from 'fs/promises'
 import { readFileSync } from 'fs'
 import { config } from './config'
 import { getToken, requireAuth } from './services/auth'
-import { initGatewayManager } from './services/gateway-bootstrap'
+import { initGatewayManager, getGatewayManagerInstance } from './services/gateway-bootstrap'
 import { bindShutdown } from './services/shutdown'
 import { setupTerminalWebSocket } from './routes/hermes/terminal'
 import { startVersionCheck } from './routes/health'
 import { registerRoutes } from './routes'
+import { setGroupChatServer } from './routes/hermes/group-chat'
+import { GroupChatServer } from './services/hermes/group-chat'
 import { logger } from './services/logger'
 
 // Injected by esbuild at build time; fallback to reading package.json in dev mode
@@ -85,6 +87,19 @@ export async function bootstrap() {
   setupTerminalWebSocket(server)
   console.log('[bootstrap] terminal websocket setup')
 
+  // Group chat Socket.IO (must be after server is created)
+  const groupChatServer = new GroupChatServer(server)
+  setGroupChatServer(groupChatServer)
+  groupChatServer.setGatewayManager(getGatewayManagerInstance())
+
+  // Catch-all: destroy upgrade requests not handled by terminal or Socket.IO
+  server.on('upgrade', (req: any, socket: any) => {
+    const url = new URL(req.url || '', `http://${req.headers.host}`)
+    if (url.pathname !== '/api/hermes/terminal' && !url.pathname.startsWith('/socket.io/')) {
+      socket.destroy()
+    }
+  })
+
   server.on('listening', () => {
     const interfaces = os.networkInterfaces()
     const localIp = Object.values(interfaces).flat().find(i => i?.family === 'IPv4' && !i?.internal)?.address || 'localhost'
@@ -93,6 +108,9 @@ export async function bootstrap() {
     console.log(`Log: ~/.hermes-web-ui/logs/server.log`)
     logger.info('Server: http://localhost:%d (LAN: http://%s:%d)', config.port, localIp, config.port)
     logger.info('Upstream: %s', config.upstream)
+
+    // Restore group chat agents after server is ready
+    groupChatServer.restoreWhenReady()
   })
 
   server.on('error', (err: any) => {
@@ -100,7 +118,7 @@ export async function bootstrap() {
     logger.error({ err }, 'Server error')
   })
 
-  bindShutdown(server)
+  bindShutdown(server, groupChatServer)
   startVersionCheck()
 }
 
diff --git a/packages/server/src/routes/hermes/group-chat.ts b/packages/server/src/routes/hermes/group-chat.ts
new file mode 100644
index 0000000..de1d360
--- /dev/null
+++ b/packages/server/src/routes/hermes/group-chat.ts
@@ -0,0 +1,270 @@
+import Router from '@koa/router'
+import type { GroupChatServer } from '../../services/hermes/group-chat'
+
+export const groupChatRoutes = new Router()
+
+let chatServer: GroupChatServer | null = null
+
+export function setGroupChatServer(server: GroupChatServer) {
+    chatServer = server
+}
+
+export function getGroupChatServer(): GroupChatServer | null {
+    return chatServer
+}
+
+function generateId(): string {
+    return Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
+}
+
+// Create room
+groupChatRoutes.post('/api/hermes/group-chat/rooms', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const { name, inviteCode, agents, compression } = ctx.request.body as {
+        name?: string
+        inviteCode?: string
+        agents?: { profile: string; name?: string; description?: string; invited?: boolean }[]
+        compression?: { triggerTokens?: number; maxHistoryTokens?: number; tailMessageCount?: number }
+    }
+    if (!name || !inviteCode) {
+        ctx.status = 400
+        ctx.body = { error: 'name and inviteCode are required' }
+        return
+    }
+
+    const roomId = generateId()
+    const storage = chatServer.getStorage()
+    storage.saveRoom(roomId, name, inviteCode, compression)
+
+    // Save agents to DB and auto-connect via Socket.IO
+    const addedAgents = []
+    for (const a of agents || []) {
+        const agentId = generateId()
+        const agent = storage.addRoomAgent(roomId, agentId, a.profile, a.name || a.profile, a.description || '', a.invited ? 1 : 0)
+        addedAgents.push(agent)
+
+        try {
+            const client = await chatServer.agentClients.createAgent({
+                profile: agent.profile,
+                name: agent.name,
+                description: agent.description,
+                invited: agent.invited,
+            })
+            await chatServer.agentClients.addAgentToRoom(roomId, client)
+        } catch (err: any) {
+            console.error(`[GroupChat] Failed to connect agent ${a.profile} to room ${roomId}: ${err.message}`)
+        }
+    }
+
+    const room = storage.getRoom(roomId)
+    ctx.body = { room, agents: addedAgents }
+})
+
+// Get room detail and messages
+groupChatRoutes.get('/api/hermes/group-chat/rooms/:roomId', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const room = chatServer.getStorage().getRoom(ctx.params.roomId)
+    if (!room) {
+        ctx.status = 404
+        ctx.body = { error: 'Room not found' }
+        return
+    }
+
+    const messages = chatServer.getStorage().getMessages(ctx.params.roomId)
+    const agents = chatServer.getStorage().getRoomAgents(ctx.params.roomId)
+    const members = chatServer.getStorage().getRoomMembers(ctx.params.roomId)
+    ctx.body = { room, messages, agents, members }
+})
+
+// List rooms
+groupChatRoutes.get('/api/hermes/group-chat/rooms', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const rooms = chatServer.getStorage().getAllRooms()
+    ctx.body = { rooms }
+})
+
+// Get room by invite code
+groupChatRoutes.get('/api/hermes/group-chat/rooms/join/:code', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const room = chatServer.getStorage().getRoomByInviteCode(ctx.params.code)
+    if (!room) {
+        ctx.status = 404
+        ctx.body = { error: 'Room not found' }
+        return
+    }
+
+    ctx.body = { room }
+})
+
+// Update room invite code
+groupChatRoutes.put('/api/hermes/group-chat/rooms/:roomId/invite-code', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const { inviteCode } = ctx.request.body as { inviteCode?: string }
+    if (!inviteCode) {
+        ctx.status = 400
+        ctx.body = { error: 'inviteCode is required' }
+        return
+    }
+
+    chatServer.getStorage().updateRoomInviteCode(ctx.params.roomId, inviteCode)
+    ctx.body = { success: true }
+})
+
+// Add agent to room
+groupChatRoutes.post('/api/hermes/group-chat/rooms/:roomId/agents', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const { profile, name, description, invited } = ctx.request.body as { profile?: string; name?: string; description?: string; invited?: boolean }
+    if (!profile) {
+        ctx.status = 400
+        ctx.body = { error: 'profile is required' }
+        return
+    }
+
+    // Prevent duplicate agent in same room
+    const existing = chatServer.getStorage().getRoomAgents(ctx.params.roomId)
+    if (existing.find(a => a.profile === profile)) {
+        ctx.status = 409
+        ctx.body = { error: 'Agent already in room' }
+        return
+    }
+
+    const agentId = generateId()
+    const agent = chatServer.getStorage().addRoomAgent(ctx.params.roomId, agentId, profile, name || profile, description || '', invited ? 1 : 0)
+
+    // Auto-connect agent via Socket.IO
+    try {
+        const client = await chatServer.agentClients.createAgent({
+            profile: agent.profile,
+            name: agent.name,
+            description: agent.description,
+            invited: agent.invited,
+        })
+        await chatServer.agentClients.addAgentToRoom(ctx.params.roomId, client)
+    } catch (err: any) {
+        console.error(`[GroupChat] Failed to connect agent ${profile} to room ${ctx.params.roomId}: ${err.message}`)
+    }
+
+    ctx.body = { agent }
+})
+
+// List agents in room
+groupChatRoutes.get('/api/hermes/group-chat/rooms/:roomId/agents', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const agents = chatServer.getStorage().getRoomAgents(ctx.params.roomId)
+    ctx.body = { agents }
+})
+
+// Remove agent from room
+groupChatRoutes.delete('/api/hermes/group-chat/rooms/:roomId/agents/:agentId', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    chatServer.getStorage().removeRoomAgent(ctx.params.agentId)
+    chatServer.agentClients.removeAgentFromRoom(ctx.params.roomId, ctx.params.agentId)
+    ctx.body = { success: true }
+})
+
+// Delete room
+groupChatRoutes.delete('/api/hermes/group-chat/rooms/:roomId', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const roomId = ctx.params.roomId
+    // Disconnect all agents in room
+    chatServer.agentClients.disconnectRoom(roomId)
+    // Delete all data
+    chatServer.getStorage().deleteRoom(roomId)
+    ctx.body = { success: true }
+})
+
+// Update room compression config
+groupChatRoutes.put('/api/hermes/group-chat/rooms/:roomId/config', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const roomId = ctx.params.roomId
+    const { triggerTokens, maxHistoryTokens, tailMessageCount } = ctx.request.body as {
+        triggerTokens?: number
+        maxHistoryTokens?: number
+        tailMessageCount?: number
+    }
+
+    chatServer.getStorage().updateRoomConfig(roomId, { triggerTokens, maxHistoryTokens, tailMessageCount })
+    const room = chatServer.getStorage().getRoom(roomId)
+    ctx.body = { room }
+})
+
+// Force compress a room's context
+groupChatRoutes.post('/api/hermes/group-chat/rooms/:roomId/compress', async (ctx) => {
+    if (!chatServer) {
+        ctx.status = 503
+        ctx.body = { error: 'Group chat not initialized' }
+        return
+    }
+
+    const roomId = ctx.params.roomId
+    if (!chatServer.getStorage().getRoom(roomId)) {
+        ctx.status = 404
+        ctx.body = { error: 'Room not found' }
+        return
+    }
+
+    const engine = chatServer.getContextEngine()
+    if (!engine) {
+        ctx.status = 503
+        ctx.body = { error: 'Context engine not available' }
+        return
+    }
+
+    try {
+        const result = await engine.forceCompress(roomId)
+        ctx.body = { success: true, summary: result }
+    } catch (err: any) {
+        ctx.status = 500
+        ctx.body = { error: err.message }
+    }
+})
diff --git a/packages/server/src/routes/hermes/terminal.ts b/packages/server/src/routes/hermes/terminal.ts
index 2a531a1..3b57cff 100644
--- a/packages/server/src/routes/hermes/terminal.ts
+++ b/packages/server/src/routes/hermes/terminal.ts
@@ -122,7 +122,6 @@ export function setupTerminalWebSocket(httpServer: HttpServer) {
   httpServer.on('upgrade', async (req, socket, head) => {
     const url = new URL(req.url || '', `http://${req.headers.host}`)
     if (url.pathname !== '/api/hermes/terminal') {
-      socket.destroy()
       return
     }
 
diff --git a/packages/server/src/routes/index.ts b/packages/server/src/routes/index.ts
index 4c43964..128b46d 100644
--- a/packages/server/src/routes/index.ts
+++ b/packages/server/src/routes/index.ts
@@ -24,6 +24,7 @@ import { fileRoutes } from './hermes/files'
 import { downloadRoutes } from './hermes/download'
 import { jobRoutes } from './hermes/jobs'
 import { proxyRoutes, proxyMiddleware } from './hermes/proxy'
+import { groupChatRoutes, setGroupChatServer } from './hermes/group-chat'
 
 /**
  * Register all routes on the Koa app.
@@ -55,6 +56,7 @@ export function registerRoutes(app: any, requireAuth: (ctx: Context, next: Next)
   app.use(nousAuthRoutes.routes())
   app.use(gatewayRoutes.routes())
   app.use(weixinRoutes.routes())
+  app.use(groupChatRoutes.routes())       // Must be before proxy
   app.use(fileRoutes.routes())              // Must be before proxy (proxy catch-all matches everything)
   app.use(downloadRoutes.routes())          // Must be before proxy
   app.use(jobRoutes.routes())               // Must be before proxy
diff --git a/packages/server/src/services/hermes/context-engine/compressor.ts b/packages/server/src/services/hermes/context-engine/compressor.ts
new file mode 100644
index 0000000..d4b4f44
--- /dev/null
+++ b/packages/server/src/services/hermes/context-engine/compressor.ts
@@ -0,0 +1,357 @@
+import type {
+    StoredMessage,
+    CompressionConfig,
+    CompressedContext,
+    BuildContextInput,
+    MessageFetcher,
+    GatewayCaller,
+    SessionCleaner,
+} from './types'
+import { DEFAULT_COMPRESSION_CONFIG } from './types'
+import { GatewaySummarizer } from './gateway-client'
+import { buildAgentInstructions, buildSummarizationSystemPrompt } from './prompt'
+
+export class ContextEngine {
+    private config: CompressionConfig
+    private messageFetcher: MessageFetcher
+    private gatewayCaller: GatewayCaller
+    /** Per-room compression lock to prevent concurrent snapshot overwrites */
+    private _compressLocks = new Map<string, Promise<void>>()
+    private _upstream = ''
+    private _apiKey: string | null = null
+
+    constructor(opts: {
+        config?: Partial<CompressionConfig>
+        messageFetcher: MessageFetcher
+        gatewayCaller?: GatewayCaller
+        sessionCleaner?: SessionCleaner
+    }) {
+        this.config = { ...DEFAULT_COMPRESSION_CONFIG, ...opts.config }
+        this.messageFetcher = opts.messageFetcher
+        this.gatewayCaller = opts.gatewayCaller || new GatewaySummarizer(this.config.summarizationTimeoutMs)
+        this.sessionCleaner = opts.sessionCleaner
+    }
+
+    private sessionCleaner?: SessionCleaner
+
+    setUpstream(upstream: string, apiKey: string | null): void {
+        this._upstream = upstream
+        this._apiKey = apiKey
+    }
+
+    /**
+     * Build context for an agent reply.
+     *
+     * Flow:
+     * 1. Read persisted snapshot (summary + lastMessageId) from SQLite
+     * 2. If snapshot exists:
+     *    a. Collect new messages after lastMessageId
+     *    b. Estimate tokens = summary + new messages
+     *    c. Under threshold → return as-is
+     *    d. Over threshold → incremental compress, update snapshot, return
+     * 3. If no snapshot:
+     *    a. Estimate tokens for all messages
+     *    b. Under threshold → return all verbatim
+     *    c. Over threshold → full compress, save snapshot, return
+     */
+    async buildContext(input: BuildContextInput): Promise<CompressedContext> {
+        // Serialize compression per room to prevent concurrent snapshot overwrites
+        const existing = this._compressLocks.get(input.roomId)
+        if (existing) {
+            await existing
+        }
+        let resolveLock!: () => void
+        const lock = new Promise<void>(r => { resolveLock = r })
+        this._compressLocks.set(input.roomId, lock)
+        try {
+            return await this._buildContextImpl(input)
+        } finally {
+            resolveLock()
+            this._compressLocks.delete(input.roomId)
+        }
+    }
+
+    private async _buildContextImpl(input: BuildContextInput): Promise<CompressedContext> {
+        const config = { ...this.config, ...input.compression }
+        const allMessages = this.messageFetcher.getMessages(input.roomId)
+        // Filter out messages newer than the current one
+        const messages = allMessages.filter(m => m.timestamp <= input.currentMessage.timestamp)
+        const total = messages.length
+
+        console.log(`[ContextEngine] buildContext START — room=${input.roomId}, agent=${input.agentName}, totalMessagesInDb=${allMessages.length}, afterFilter=${total}`)
+
+        const instructions = buildAgentInstructions({
+            agentName: input.agentName,
+            roomName: input.roomName,
+            agentDescription: input.agentDescription,
+            memberNames: input.memberNames,
+            members: input.members,
+        })
+
+        const meta: CompressedContext['meta'] = {
+            totalMessages: total,
+            verbatimCount: 0,
+            hadSnapshot: false,
+            compressed: false,
+            summaryTokenEstimate: 0,
+        }
+
+        const snapshot = this.messageFetcher.getContextSnapshot(input.roomId)
+        console.log(`[ContextEngine] snapshot=${snapshot ? `EXISTS (lastMsgId=${snapshot.lastMessageId}, summaryLen=${snapshot.summary.length})` : 'NONE'}`)
+
+        // ── Path A: Snapshot exists — incremental ────────────
+        if (snapshot) {
+            meta.hadSnapshot = true
+
+            // Find the position of lastMessageId in messages
+            const snapshotIdx = messages.findIndex(m => m.id === snapshot.lastMessageId)
+            // Collect messages after the snapshot position
+            const newMessages = snapshotIdx >= 0
+                ? messages.slice(snapshotIdx + 1)
+                : messages.filter(m => m.timestamp > snapshot.lastMessageTimestamp)
+
+            const summaryTokens = this.countTokens(snapshot.summary)
+            const newTokens = this.estimateTokensFromMessages(newMessages)
+            const totalTokens = summaryTokens + newTokens
+
+            meta.verbatimCount = newMessages.length
+            meta.summaryTokenEstimate = summaryTokens
+
+            console.log(`[ContextEngine] [Path A] snapshotIdx=${snapshotIdx}, newMessages=${newMessages.length}, summaryTokens=~${summaryTokens}, newTokens=~${newTokens}, totalTokens=~${totalTokens}, threshold=${config.triggerTokens}`)
+            console.log(`[ContextEngine] [Path A] EXISTING SUMMARY (${snapshot.summary.length} chars):`, snapshot.summary.slice(0, 300))
+            if (newMessages.length > 0) {
+                console.log(`[ContextEngine] [Path A] NEW MESSAGES (${newMessages.length}):`, newMessages.map(m => `[${m.senderName}]: ${m.content.slice(0, 80)}`).join(' | '))
+            }
+
+            // Under threshold — return summary + new messages directly
+            if (totalTokens <= config.triggerTokens) {
+                console.log(`[ContextEngine] [Path A] UNDER threshold — return summary + ${newMessages.length} verbatim msgs directly`)
+                const history = this.buildHistory(snapshot.summary, newMessages, input.agentSocketId)
+                this.logHistory('Path A (no compress)', history)
+                return { conversationHistory: history, instructions, meta }
+            }
+
+            // Over threshold — incremental compress
+            console.log(`[ContextEngine] [Path A] OVER threshold — starting INCREMENTAL compression of ${newMessages.length} msgs...`)
+            console.log(`[ContextEngine] [Path A] CONTEXT BEFORE COMPRESSION: summary(${snapshot.summary.length} chars) + ${newMessages.length} new msgs`)
+            meta.compressed = true
+
+            const t0 = Date.now()
+            const result = await this.summarize(
+                input.roomId,
+                newMessages,
+                input.upstream,
+                input.apiKey,
+                snapshot.summary,
+            )
+            const elapsed = Date.now() - t0
+
+            if (result.summary) {
+                const lastMsg = newMessages[newMessages.length - 1]
+                this.messageFetcher.saveContextSnapshot(input.roomId, result.summary, lastMsg.id, lastMsg.timestamp)
+
+                meta.summaryTokenEstimate = this.countTokens(result.summary)
+                console.log(`[ContextEngine] [Path A] incremental compression DONE in ${elapsed}ms, newSummaryLen=${result.summary.length}, newLastMsgId=${lastMsg.id}`)
+                console.log(`[ContextEngine] [Path A] NEW SUMMARY (${result.summary.length} chars):`, result.summary.slice(0, 300))
+                const history = this.buildHistory(result.summary, newMessages, input.agentSocketId)
+                this.logHistory('Path A (after incremental compress)', history)
+                if (result.sessionId) this.sessionCleaner?.(result.sessionId)
+                return { conversationHistory: history, instructions, meta }
+            }
+
+            // Compression failed — degrade
+            console.warn(`[ContextEngine] [Path A] incremental compression FAILED (${elapsed}ms) — degrading to summary + trimmed verbatim`)
+            const history = this.buildHistory(snapshot.summary, newMessages, input.agentSocketId)
+            this.trimToBudget(history, summaryTokens, config.maxHistoryTokens)
+            return { conversationHistory: history, instructions, meta }
+        }
+
+        // ── Path B: No snapshot — full context ───────────────
+        const totalTokens = this.estimateTokensFromMessages(messages)
+        meta.verbatimCount = total
+
+        console.log(`[ContextEngine] [Path B] no snapshot, totalMessages=${total}, totalTokens=~${totalTokens}, threshold=${config.triggerTokens}`)
+
+        // Under threshold — pass all messages verbatim
+        if (totalTokens <= config.triggerTokens) {
+            console.log(`[ContextEngine] [Path B] UNDER threshold — return all ${total} msgs verbatim`)
+            const history = messages.map(m => this.mapToHistory(m, input.agentSocketId))
+            this.logHistory('Path B (no compress)', history)
+            return { conversationHistory: history, instructions, meta }
+        }
+
+        // Over threshold — full compress
+        console.log(`[ContextEngine] [Path B] OVER threshold — starting FULL compression of ${total} msgs...`)
+        console.log(`[ContextEngine] [Path B] CONTEXT BEFORE COMPRESSION: ${total} msgs, ~${totalTokens} tokens`)
+        meta.compressed = true
+
+        const t0 = Date.now()
+        const result = await this.summarize(
+            input.roomId,
+            messages,
+            input.upstream,
+            input.apiKey,
+        )
+        const elapsed = Date.now() - t0
+
+        if (result.summary) {
+            // Keep recent tail messages verbatim, compress the rest
+            const { tailMessageCount } = config
+            const toCompress = messages.length > tailMessageCount ? messages.slice(0, -tailMessageCount) : messages
+            const tail = messages.length > tailMessageCount ? messages.slice(-tailMessageCount) : []
+            const lastCompressedMsg = toCompress[toCompress.length - 1]
+
+            this.messageFetcher.saveContextSnapshot(input.roomId, result.summary, lastCompressedMsg.id, lastCompressedMsg.timestamp)
+
+            meta.summaryTokenEstimate = this.countTokens(result.summary)
+            console.log(`[ContextEngine] [Path B] full compression DONE in ${elapsed}ms, summaryLen=${result.summary.length}, compressed=${toCompress.length} msgs, keptTail=${tail.length} msgs, savedLastMsgId=${lastCompressedMsg.id}`)
+            console.log(`[ContextEngine] [Path B] COMPRESSED SUMMARY (${result.summary.length} chars):`, result.summary.slice(0, 300))
+            const history = this.buildHistory(result.summary, tail, input.agentSocketId)
+            this.logHistory('Path B (after full compress)', history)
+            if (result.sessionId) this.sessionCleaner?.(result.sessionId)
+            return { conversationHistory: history, instructions, meta }
+        }
+
+        // Compression failed — degrade
+        console.warn(`[ContextEngine] [Path B] full compression FAILED (${elapsed}ms) — degrading to trimmed verbatim`)
+        const history = messages.map(m => this.mapToHistory(m, input.agentSocketId))
+        this.trimToBudget(history, 0, config.maxHistoryTokens)
+        meta.verbatimCount = history.length
+        return { conversationHistory: history, instructions, meta }
+    }
+
+    invalidateRoom(roomId: string): void {
+        this.messageFetcher.deleteContextSnapshot(roomId)
+    }
+
+    /**
+     * Force compress all messages in a room (full compression).
+     * Used when user manually triggers compression.
+     */
+    async forceCompress(roomId: string): Promise<string> {
+        const allMessages = this.messageFetcher.getMessages(roomId)
+        if (allMessages.length === 0) return ''
+
+        const config = { ...this.config }
+        console.log(`[ContextEngine] forceCompress room=${roomId}, messages=${allMessages.length}`)
+
+        const t0 = Date.now()
+        const result = await this.summarize(roomId, allMessages, this._upstream, this._apiKey)
+        const elapsed = Date.now() - t0
+
+        if (result.summary) {
+            const { tailMessageCount } = config
+            const toCompress = allMessages.length > tailMessageCount ? allMessages.slice(0, -tailMessageCount) : allMessages
+            const tail = allMessages.length > tailMessageCount ? allMessages.slice(-tailMessageCount) : []
+            const lastCompressedMsg = toCompress[toCompress.length - 1]
+
+            this.messageFetcher.saveContextSnapshot(roomId, result.summary, lastCompressedMsg.id, lastCompressedMsg.timestamp)
+            console.log(`[ContextEngine] forceCompress DONE in ${elapsed}ms`)
+            if (result.sessionId) this.sessionCleaner?.(result.sessionId)
+            return result.summary
+        }
+
+        throw new Error('Compression failed')
+    }
+
+    // ─── Private ─────────────────────────────────────────────
+
+    /**
+     * Build history array: optional summary prefix + verbatim messages.
+     */
+    private buildHistory(
+        summary: string,
+        messages: StoredMessage[],
+        agentSocketId: string,
+    ): Array<{ role: 'user' | 'assistant'; content: string }> {
+        const history: Array<{ role: 'user' | 'assistant'; content: string }> = []
+
+        if (summary) {
+            history.push(
+                { role: 'user', content: '[Previous conversation summary]\n' + summary },
+                { role: 'assistant', content: 'I have reviewed the conversation history and understand the context.' },
+            )
+        }
+
+        history.push(...messages.map(m => this.mapToHistory(m, agentSocketId)))
+        return history
+    }
+
+    /**
+     * Summarize messages. If previousSummary is provided, do incremental update.
+     */
+    private async summarize(
+        roomId: string,
+        messages: StoredMessage[],
+        upstream: string,
+        apiKey: string | null,
+        previousSummary?: string,
+    ): Promise<{ summary: string | null; sessionId: string | null }> {
+        if (messages.length === 0 && !previousSummary) return { summary: null, sessionId: null }
+
+        try {
+            const result = await this.gatewayCaller.summarize(
+                upstream,
+                apiKey,
+                buildSummarizationSystemPrompt(),
+                messages,
+                previousSummary,
+            )
+            return { summary: result.summary, sessionId: result.sessionId }
+        } catch (err: any) {
+            console.warn(`[ContextEngine] Summarization failed for room ${roomId}: ${err.message}`)
+            return { summary: null, sessionId: null }
+        } finally {
+            // Session cleanup handled here if sessionCleaner is provided
+        }
+    }
+
+    private mapToHistory(
+        msg: StoredMessage,
+        agentSocketId: string,
+    ): { role: 'user' | 'assistant'; content: string } {
+        if (msg.senderId === agentSocketId) {
+            return { role: 'assistant', content: msg.content }
+        }
+        return { role: 'user', content: `[${msg.senderName}]: ${msg.content}` }
+    }
+
+    private trimToBudget(
+        history: Array<{ role: 'user' | 'assistant'; content: string }>,
+        summaryTokens: number,
+        maxTokens: number,
+    ): void {
+        let totalTokens = summaryTokens + this.estimateTokens(history)
+        while (totalTokens > maxTokens && history.length > 0) {
+            history.pop()
+            totalTokens = summaryTokens + this.estimateTokens(history)
+        }
+    }
+
+    private estimateTokens(history: Array<{ role: string; content: string }>): number {
+        const text = history.map(m => m.content).join('')
+        return this.countTokens(text)
+    }
+
+    private estimateTokensFromMessages(messages: StoredMessage[]): number {
+        const text = messages.map(m => m.content + m.senderName).join('')
+        return this.countTokens(text)
+    }
+
+    /** Estimate tokens distinguishing CJK (~1.5 tok/char) from Latin (~0.25 tok/char) */
+    private countTokens(text: string): number {
+        const cjk = (text.match(/[\u2e80-\u9fff\uac00-\ud7af\u3000-\u303f\uff00-\uffef]/g) || []).length
+        const other = text.length - cjk
+        return Math.ceil(cjk * 1.5 + other / 4)
+    }
+
+    /** Log assembled history for debugging */
+    private logHistory(label: string, history: Array<{ role: string; content: string }>): void {
+        const totalTokens = this.estimateTokens(history)
+        console.log(`[ContextEngine] ASSEMBLED HISTORY (${label}): ${history.length} entries, ~${totalTokens} tokens`)
+        for (const entry of history) {
+            const preview = entry.content.length > 150 ? entry.content.slice(0, 150) + '...' : entry.content
+            console.log(`  [${entry.role}] ${preview}`)
+        }
+    }
+}
diff --git a/packages/server/src/services/hermes/context-engine/gateway-client.ts b/packages/server/src/services/hermes/context-engine/gateway-client.ts
new file mode 100644
index 0000000..decfb55
--- /dev/null
+++ b/packages/server/src/services/hermes/context-engine/gateway-client.ts
@@ -0,0 +1,117 @@
+import { EventSource } from 'eventsource'
+import type { StoredMessage, GatewayCaller } from './types'
+import {
+    buildSummarizationSystemPrompt,
+    buildFullSummaryPrompt,
+    buildIncrementalUpdatePrompt,
+} from './prompt'
+
+/**
+ * Calls Hermes /v1/runs to produce LLM-generated summaries.
+ * Uses non-streaming EventSource to wait for run.completed.
+ */
+export class GatewaySummarizer implements GatewayCaller {
+    private timeoutMs: number
+
+    constructor(timeoutMs = 30_000) {
+        this.timeoutMs = timeoutMs
+    }
+
+    async summarize(
+        upstream: string,
+        apiKey: string | null,
+        systemPrompt: string,
+        messages: StoredMessage[],
+        previousSummary?: string,
+    ): Promise<{ summary: string; sessionId: string }> {
+        // Build conversation_history from messages
+        const history: Array<{ role: string; content: string }> = messages.map(m => ({
+            role: 'user',
+            content: `[${m.senderName}]: ${m.content}`,
+        }))
+
+        // Inject previous summary for incremental update
+        if (previousSummary) {
+            history.unshift(
+                { role: 'user', content: `[Previous summary]\n${previousSummary}` },
+                { role: 'assistant', content: 'Understood, I will update the summary.' },
+            )
+        }
+
+        const userPrompt = previousSummary
+            ? buildIncrementalUpdatePrompt()
+            : buildFullSummaryPrompt()
+
+        const sessionId = Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
+
+        // POST /v1/runs
+        const res = await fetch(`${upstream}/v1/runs`, {
+            method: 'POST',
+            headers: {
+                'Content-Type': 'application/json',
+                ...(apiKey ? { Authorization: `Bearer ${apiKey}` } : {}),
+            },
+            body: JSON.stringify({
+                input: userPrompt,
+                instructions: systemPrompt || buildSummarizationSystemPrompt(),
+                conversation_history: history,
+                session_id: sessionId,
+            }),
+            signal: AbortSignal.timeout(this.timeoutMs),
+        })
+
+        if (!res.ok) {
+            throw new Error(`Summarization run failed: ${res.status}`)
+        }
+
+        const { run_id } = await res.json() as { run_id: string }
+
+        try {
+            const output = await this.pollForResult(upstream, apiKey, run_id)
+            return { summary: output, sessionId }
+        } finally {
+            // Note: session cleanup is handled by the caller (compressor.ts)
+        }
+    }
+
+    private pollForResult(upstream: string, apiKey: string | null, runId: string): Promise<string> {
+        return new Promise<string>((resolve, reject) => {
+            const timer = setTimeout(() => {
+                source.close()
+                reject(new Error('Summarization timed out'))
+            }, this.timeoutMs)
+
+            const eventsUrl = new URL(`${upstream}/v1/runs/${runId}/events`)
+            if (apiKey) eventsUrl.searchParams.set('token', apiKey)
+
+            const source = new EventSource(eventsUrl.toString())
+
+            source.onmessage = (event: MessageEvent) => {
+                try {
+                    const parsed = JSON.parse(event.data)
+                    if (parsed.event === 'run.completed') {
+                        clearTimeout(timer)
+                        source.close()
+                        const output = parsed.output
+                        if (!output || typeof output !== 'string' || output.trim() === '') {
+                            reject(new Error('Empty summarization response'))
+                            return
+                        }
+                        resolve(output.trim())
+                    } else if (parsed.event === 'run.failed') {
+                        clearTimeout(timer)
+                        source.close()
+                        reject(new Error(parsed.error || 'Summarization run failed'))
+                    }
+                } catch { /* ignore parse errors for non-JSON events */ }
+            }
+
+            source.onerror = () => {
+                clearTimeout(timer)
+                source.close()
+                reject(new Error('Summarization SSE connection error'))
+            }
+        })
+    }
+
+}
diff --git a/packages/server/src/services/hermes/context-engine/index.ts b/packages/server/src/services/hermes/context-engine/index.ts
new file mode 100644
index 0000000..9817bd6
--- /dev/null
+++ b/packages/server/src/services/hermes/context-engine/index.ts
@@ -0,0 +1,13 @@
+export { ContextEngine } from './compressor'
+export { GatewaySummarizer } from './gateway-client'
+export { buildAgentInstructions, buildSummarizationSystemPrompt, buildFullSummaryPrompt, buildIncrementalUpdatePrompt } from './prompt'
+export { DEFAULT_COMPRESSION_CONFIG } from './types'
+export type {
+    StoredMessage,
+    CompressionConfig,
+    CompressedContext,
+    ContextSnapshot,
+    MessageFetcher,
+    GatewayCaller,
+    BuildContextInput,
+} from './types'
diff --git a/packages/server/src/services/hermes/context-engine/prompt.ts b/packages/server/src/services/hermes/context-engine/prompt.ts
new file mode 100644
index 0000000..fc0afb5
--- /dev/null
+++ b/packages/server/src/services/hermes/context-engine/prompt.ts
@@ -0,0 +1,82 @@
+// ─── Agent Identity Instructions ────────────────────────────
+
+import type { MemberInfo } from './types'
+
+interface AgentInstructionsParams {
+    agentName: string
+    roomName: string
+    agentDescription: string
+    memberNames: string[]
+    members: MemberInfo[]
+}
+
+export function buildAgentInstructions(params: AgentInstructionsParams): string {
+    let memberSection: string
+    if (params.members.length > 0) {
+        memberSection = params.members
+            .map(m => m.description ? `- ${m.name}: ${m.description}` : `- ${m.name}`)
+            .join('\n')
+    } else if (params.memberNames.length > 0) {
+        memberSection = params.memberNames.map(n => `- ${n}`).join('\n')
+    } else {
+        memberSection = '- 未知'
+    }
+
+    return `你是"${params.agentName}"，群聊房间"${params.roomName}"中的 AI 助手。
+
+你的角色：${params.agentDescription}
+
+当前房间成员：
+${memberSection}
+
+规则：
+- 有人用 @${params.agentName} 提及你时才需要回复，重点回应提及你的人。
+- 回答简洁、对群聊有帮助。
+- 不要假装是人类，需要时明确表明自己是 AI。
+- 对话历史中包含多个人的消息，每条消息前标有发送者名字。
+- 对话开头可能包含之前的对话摘要，用于提供更早的上下文。
+- 回复最新一条提及你的消息。
+- 如果需要其他 agent 协作或明确回复某个人，使用 @名字 来提及对方。
+- 自行判断对话是否已经结束——如果问题已解决、达成共识、或对方只是陈述不需要回复，则不要再 @任何人，直接结束回复，避免产生无意义的循环对话。`
+}
+
+// ─── Summarization Prompts ─────────────────────────────────
+
+export function buildSummarizationSystemPrompt(): string {
+    return `你是一个群聊对话的摘要助手。请创建一份结构化摘要，帮助 AI 助手快速理解完整的对话上下文并智能回复。
+
+使用以下格式：
+
+当前话题：
+- 现在在聊什么，目标是什么
+
+已知结论：
+- 已达成哪些共识，哪些问题已经回答过
+
+待回复消息：
+- 还剩谁的问题没回，下一步要做什么
+
+关键人物：
+- 人名、角色、引用关系
+
+重要上下文：
+- 不要丢时间线和立场变化
+- 少写废话，多保留"可行动信息"
+- 重点保留：谁说了什么、结论是什么、下一步是什么
+- 关键的 URL、代码片段、错误信息、约束条件
+
+规则：
+- 基于事实，不要编造信息。
+- 保持简洁（500 字以内）。
+- 聚焦于帮助 AI 回复下一条消息的可行动信息。
+- 使用与对话相同的语言。
+- 不要回复对话内容，只输出摘要。`
+}
+
+export function buildFullSummaryPrompt(): string {
+    return '请对上方对话创建一份简洁的摘要。只输出摘要内容。'
+}
+
+export function buildIncrementalUpdatePrompt(): string {
+    return '对话自上次摘要后有了新的内容。请更新摘要，整合新消息。保持相同格式，更新所有部分。只输出更新后的摘要。'
+}
diff --git a/packages/server/src/services/hermes/context-engine/summary-cache.ts b/packages/server/src/services/hermes/context-engine/summary-cache.ts
new file mode 100644
index 0000000..5adebdd
--- /dev/null
+++ b/packages/server/src/services/hermes/context-engine/summary-cache.ts
@@ -0,0 +1,49 @@
+import type { SummaryCacheEntry } from './types'
+
+const MAX_ENTRIES = 200
+
+export class SummaryCache {
+    private cache = new Map<string, SummaryCacheEntry>()
+    private ttlMs: number
+
+    constructor(ttlMs = 120_000) {
+        this.ttlMs = ttlMs
+    }
+
+    get(roomId: string): SummaryCacheEntry | undefined {
+        const entry = this.cache.get(roomId)
+        if (!entry) return undefined
+        if (Date.now() - entry.createdAt >= this.ttlMs) {
+            this.cache.delete(roomId)
+            return undefined
+        }
+        return entry
+    }
+
+    set(roomId: string, entry: SummaryCacheEntry): void {
+        if (this.cache.size >= MAX_ENTRIES) {
+            let oldestKey = ''
+            let oldestTime = Infinity
+            for (const [k, v] of this.cache) {
+                if (v.createdAt < oldestTime) {
+                    oldestTime = v.createdAt
+                    oldestKey = k
+                }
+            }
+            if (oldestKey) this.cache.delete(oldestKey)
+        }
+        this.cache.set(roomId, entry)
+    }
+
+    invalidate(roomId: string): void {
+        this.cache.delete(roomId)
+    }
+
+    clear(): void {
+        this.cache.clear()
+    }
+
+    get size(): number {
+        return this.cache.size
+    }
+}
diff --git a/packages/server/src/services/hermes/context-engine/types.ts b/packages/server/src/services/hermes/context-engine/types.ts
new file mode 100644
index 0000000..14bee1e
--- /dev/null
+++ b/packages/server/src/services/hermes/context-engine/types.ts
@@ -0,0 +1,111 @@
+// ─── Message Types ──────────────────────────────────────────
+
+/** Raw message from SQLite messages table */
+export interface StoredMessage {
+    id: string
+    roomId: string
+    senderId: string
+    senderName: string
+    content: string
+    timestamp: number
+}
+
+// ─── Compression Config ────────────────────────────────────
+
+export interface CompressionConfig {
+    /** Token threshold to trigger compression (estimate all messages) */
+    triggerTokens: number
+    /** Max tokens for the final compressed context sent to LLM */
+    maxHistoryTokens: number
+    /** Number of recent messages to keep verbatim after compression */
+    tailMessageCount: number
+    /** Characters per token for estimation */
+    charsPerToken: number
+    /** Timeout for summarization LLM call in ms */
+    summarizationTimeoutMs: number
+}
+
+export const DEFAULT_COMPRESSION_CONFIG: CompressionConfig = {
+    triggerTokens: 100_000,
+    maxHistoryTokens: 32_000,
+    tailMessageCount: 20,
+    charsPerToken: 4,
+    summarizationTimeoutMs: 30_000,
+}
+
+// ─── Compression Output ────────────────────────────────────
+
+export interface CompressedContext {
+    conversationHistory: Array<{ role: 'user' | 'assistant'; content: string }>
+    instructions: string
+    meta: {
+        totalMessages: number
+        verbatimCount: number
+        hadSnapshot: boolean
+        compressed: boolean
+        summaryTokenEstimate: number
+    }
+}
+
+// ─── Context Snapshot (persisted in SQLite) ────────────────
+
+export interface ContextSnapshot {
+    roomId: string
+    summary: string
+    lastMessageId: string
+    lastMessageTimestamp: number
+    updatedAt: number
+}
+
+// ─── Summary Cache ──────────────────────────────────────────
+
+export interface SummaryCacheEntry {
+    summary: string
+    lastMessageId: string
+    lastMessageTimestamp: number
+    createdAt: number
+}
+
+// ─── Dependency Injection ──────────────────────────────────
+
+export interface MessageFetcher {
+    getMessages(roomId: string, limit?: number): StoredMessage[]
+    getContextSnapshot(roomId: string): ContextSnapshot | null
+    saveContextSnapshot(roomId: string, summary: string, lastMessageId: string, lastMessageTimestamp: number): void
+    deleteContextSnapshot(roomId: string): void
+}
+
+export interface GatewayCaller {
+    summarize(
+        upstream: string,
+        apiKey: string | null,
+        systemPrompt: string,
+        messages: StoredMessage[],
+        previousSummary?: string,
+    ): Promise<{ summary: string; sessionId: string }>
+}
+
+export type SessionCleaner = (sessionId: string) => void
+
+// ─── Build Context Input ───────────────────────────────────
+
+export interface MemberInfo {
+    userId: string
+    name: string
+    description: string
+}
+
+export interface BuildContextInput {
+    roomId: string
+    agentId: string
+    agentName: string
+    agentDescription: string
+    agentSocketId: string
+    roomName: string
+    memberNames: string[]
+    members: MemberInfo[]
+    upstream: string
+    apiKey: string | null
+    currentMessage: StoredMessage
+    compression?: Partial<CompressionConfig>
+}
diff --git a/packages/server/src/services/hermes/group-chat/agent-clients.ts b/packages/server/src/services/hermes/group-chat/agent-clients.ts
new file mode 100644
index 0000000..9d37c84
--- /dev/null
+++ b/packages/server/src/services/hermes/group-chat/agent-clients.ts
@@ -0,0 +1,669 @@
+import { io, Socket } from 'socket.io-client'
+import { EventSource } from 'eventsource'
+import { getToken } from '../../../services/auth'
+import type { GatewayManager } from '../gateway-manager'
+import { deleteSession as hermesDeleteSession } from '../hermes-cli'
+import { getActiveProfileName } from '../hermes-profile'
+import { logger } from '../../../services/logger'
+
+// ─── Types ────────────────────────────────────────────────────
+
+interface AgentConfig {
+    profile: string
+    name: string
+    description: string
+    invited: number
+}
+
+interface MessageData {
+    id: string
+    roomId: string
+    senderId: string
+    senderName: string
+    content: string
+    timestamp: number
+}
+
+interface MemberData {
+    id: string
+    name: string
+    joinedAt: number
+}
+
+interface JoinResult {
+    roomId: string
+    roomName: string
+    members: MemberData[]
+    messages: MessageData[]
+    rooms: string[]
+}
+
+export interface AgentEventHandler {
+    onMessage?: (data: { roomId: string; msg: MessageData }) => void
+    onTyping?: (data: { roomId: string; userId: string; userName: string }) => void
+    onStopTyping?: (data: { roomId: string; userId: string; userName: string }) => void
+    onMemberJoined?: (data: { roomId: string; memberId: string; memberName: string; members: MemberData[] }) => void
+    onMemberLeft?: (data: { roomId: string; memberId: string; memberName: string; members: MemberData[] }) => void
+}
+
+// ─── Agent Client (single connection) ─────────────────────────
+
+class AgentClient {
+    readonly agentId: string
+    readonly profile: string
+    readonly name: string
+    readonly description: string
+    private socket: Socket | null = null
+    private joinedRooms = new Set<string>()
+    private handlers: AgentEventHandler
+    private _reconnecting = false
+    private gatewayManager: GatewayManager | null = null
+    private contextEngine: any = null
+    private storage: any = null
+
+    constructor(config: AgentConfig, handlers: AgentEventHandler = {}) {
+        this.agentId = Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
+        this.profile = config.profile
+        this.name = config.name
+        this.description = config.description
+        this.handlers = handlers
+    }
+
+    get connected(): boolean {
+        return this.socket?.connected ?? false
+    }
+
+    get id(): string | undefined {
+        return this.socket?.id
+    }
+
+    setGatewayManager(manager: GatewayManager): void {
+        this.gatewayManager = manager
+    }
+
+    setContextEngine(engine: any): void {
+        this.contextEngine = engine
+    }
+
+    setStorage(storage: any): void {
+        this.storage = storage
+    }
+
+    async connect(port = 8648): Promise<void> {
+        const token = await getToken()
+
+        this.socket = io(`http://127.0.0.1:${port}/group-chat`, {
+            auth: {
+                token: token || undefined,
+                name: this.name,
+            },
+            transports: ['websocket'],
+            reconnection: true,
+            reconnectionAttempts: Infinity,
+            reconnectionDelay: 1000,
+            reconnectionDelayMax: 30000,
+        })
+
+        this.bindEvents()
+
+        return new Promise((resolve, reject) => {
+            const timeout = setTimeout(() => reject(new Error('Connection timeout')), 10000)
+
+            this.socket!.on('connect', () => {
+                clearTimeout(timeout)
+                logger.debug(`[AgentClient] ${this.name} connected, socket id: ${this.socket!.id}`)
+                resolve()
+            })
+
+            this.socket!.on('connect_error', (err) => {
+                clearTimeout(timeout)
+                logger.error(err, `[AgentClient] ${this.name} connect_error`)
+                reject(err)
+            })
+        })
+    }
+
+    disconnect(): void {
+        if (this.socket) {
+            this.socket.disconnect()
+            this.socket = null
+            this.joinedRooms.clear()
+        }
+    }
+
+    async joinRoom(roomId: string): Promise<JoinResult> {
+        this.ensureConnected()
+        return new Promise((resolve, reject) => {
+            this.socket!.emit('join', { roomId }, (res: JoinResult | { error: string }) => {
+                if ('error' in res) {
+                    reject(new Error(res.error))
+                } else {
+                    this.joinedRooms.add(roomId)
+                    resolve(res)
+                }
+            })
+        })
+    }
+
+    sendMessage(roomId: string, content: string): Promise<string> {
+        this.ensureConnected()
+        return new Promise((resolve, reject) => {
+            this.socket!.emit('message', { roomId, content }, (res: { id?: string; error?: string }) => {
+                if (res.error) {
+                    reject(new Error(res.error))
+                } else {
+                    resolve(res.id!)
+                }
+            })
+        })
+    }
+
+    startTyping(roomId: string): void {
+        this.ensureConnected()
+        this.socket!.emit('typing', { roomId })
+    }
+
+    stopTyping(roomId: string): void {
+        this.ensureConnected()
+        this.socket!.emit('stop_typing', { roomId })
+    }
+
+    emitContextStatus(roomId: string, status: 'compressing' | 'replying' | 'ready'): void {
+        this.ensureConnected()
+        this.socket!.emit('context_status', { roomId, agentName: this.name, status })
+    }
+
+    getJoinedRooms(): string[] {
+        return Array.from(this.joinedRooms)
+    }
+
+    private ensureConnected(): void {
+        if (!this.socket?.connected) {
+            throw new Error(`Agent "${this.name}" is not connected`)
+        }
+    }
+
+    private async deleteSession(sessionId: string): Promise<void> {
+        try {
+            const sessionProfile = this.storage?.getSessionProfile?.(sessionId)
+            const currentProfile = getActiveProfileName()
+
+            if (sessionProfile && sessionProfile.profile_name !== currentProfile) {
+                // Cross-profile: enqueue deferred delete, don't switch profile
+                this.storage?.enqueuePendingSessionDelete?.(sessionId, sessionProfile.profile_name)
+                logger.info(`[AgentClients] ${this.name}: cross-profile deferred delete session ${sessionId} (session=${sessionProfile.profile_name}, active=${currentProfile})`)
+                return
+            }
+
+            // Same profile or no mapping: delete directly
+            const ok = await hermesDeleteSession(sessionId)
+            if (ok) {
+                this.storage?.deleteSessionProfile?.(sessionId)
+            }
+            logger.debug(`[AgentClients] ${this.name}: delete session ${sessionId} (profile=${this.profile}) → ${ok ? 'ok' : 'failed'}`)
+        } catch (err: any) {
+            logger.warn(`[AgentClients] ${this.name}: failed to delete session ${sessionId}: ${err.message}`)
+        }
+    }
+
+    // ─── Hermes Gateway Integration ────────────────────────────
+
+    /**
+     * Handle an @mention from the server side.
+     * Called by AgentClients.processMentions() — no socket round-trip needed.
+     * onStatus is called to report context compression progress.
+     */
+    async replyToMention(
+        roomId: string,
+        msg: { content: string; senderName: string; senderId: string; timestamp: number },
+        onStatus?: (status: 'compressing' | 'replying' | 'ready') => void,
+    ): Promise<void> {
+        logger.debug(`[AgentClients] ${this.name} mentioned by ${msg.senderName}: "${msg.content.slice(0, 50)}"`)
+        if (!this.gatewayManager) {
+            logger.debug(`[AgentClients] ${this.name}: gatewayManager is null, skipping`)
+            return
+        }
+
+        const upstream = this.gatewayManager.getUpstream(this.profile)
+        const apiKey = this.gatewayManager.getApiKey(this.profile)
+        logger.debug(`[AgentClients] ${this.name}: upstream=${upstream}, profile=${this.profile}`)
+        if (!upstream) {
+            logger.error(`[AgentClients] ${this.name}: no gateway upstream for profile "${this.profile}"`)
+            return
+        }
+
+        const sessionId = Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
+
+        try {
+            // Notify room that agent is typing
+            this.startTyping(roomId)
+
+            // Build compressed context if context engine is available
+            let conversationHistory: Array<{ role: string; content: string }> = []
+            let instructions: string | undefined
+
+            if (this.contextEngine && this.storage) {
+                try {
+                    logger.debug(`[AgentClients] ${this.name}: building context...`)
+                    onStatus?.('compressing')
+                    // Get room members with descriptions for context
+                    const roomMembers: Array<{ userId: string; name: string; description: string }> = this.storage.getRoomMembers(roomId) || []
+                    const memberNames = roomMembers.map((m: any) => m.name)
+                    const members = roomMembers.map((m: any) => ({ userId: m.userId, name: m.name, description: m.description }))
+
+                    // Get room compression config
+                    const roomInfo = this.storage.getRoom(roomId)
+                    const compression = roomInfo ? {
+                        triggerTokens: roomInfo.triggerTokens,
+                        maxHistoryTokens: roomInfo.maxHistoryTokens,
+                        tailMessageCount: roomInfo.tailMessageCount,
+                    } : undefined
+
+                    const ctx = await this.contextEngine.buildContext({
+                        roomId,
+                        agentId: this.agentId,
+                        agentName: this.name,
+                        agentDescription: this.description,
+                        agentSocketId: this.socket?.id || '',
+                        roomName: roomId,
+                        memberNames,
+                        members,
+                        upstream,
+                        apiKey,
+                        currentMessage: msg,
+                        compression,
+                    })
+                    conversationHistory = ctx.conversationHistory
+                    instructions = ctx.instructions
+                    logger.debug(`[AgentClients] ${this.name}: context built — historyLen=${conversationHistory.length}, meta=%j`, ctx.meta)
+                    onStatus?.('replying')
+                } catch (err: any) {
+                    logger.warn(`[AgentClients] ${this.name}: context engine failed: ${err.message}`)
+                    onStatus?.('replying')
+                    // Degrade: continue without context
+                }
+            }
+
+            // Strip @mention from input — agent already knows it was mentioned
+            const input = msg.content.replace(new RegExp(`@${this.name.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\s*`, 'gi'), '').trim() || msg.content
+
+            // Start a run on Hermes gateway
+            const runRes = await fetch(`${upstream}/v1/runs`, {
+                method: 'POST',
+                headers: {
+                    'Content-Type': 'application/json',
+                    ...(apiKey ? { Authorization: `Bearer ${apiKey}` } : {}),
+                },
+                body: JSON.stringify({
+                    input,
+                    session_id: sessionId,
+                    ...(conversationHistory.length > 0 ? { conversation_history: conversationHistory } : {}),
+                    ...(instructions ? { instructions } : {}),
+                }),
+                signal: AbortSignal.timeout(120000),
+            })
+
+            if (!runRes.ok) {
+                const text = await runRes.text().catch(() => '')
+                logger.error(`[AgentClients] ${this.name}: gateway run failed (${runRes.status}): ${text}`)
+                this.stopTyping(roomId)
+                return
+            }
+
+            const runData = await runRes.json() as any
+            const run_id = runData.run_id
+            logger.debug(`[AgentClients] ${this.name}: run started, response=%j`, runData)
+            if (!run_id) {
+                logger.error(`[AgentClients] ${this.name}: no run_id in response`)
+                this.stopTyping(roomId)
+                return
+            }
+
+            // Save session-to-profile mapping after gateway confirms the run
+            const actualSessionId = runData.session_id || sessionId
+            if (!this.storage) {
+                logger.warn(`[AgentClients] ${this.name}: storage is null, cannot save session profile for ${actualSessionId}`)
+            } else {
+                this.storage.saveSessionProfile(actualSessionId, roomId, this.agentId, this.profile)
+                logger.debug(`[AgentClients] ${this.name}: saved session profile ${actualSessionId} → profile=${this.profile}`)
+            }
+
+            // Stream events from Hermes
+            const eventsUrl = new URL(`${upstream}/v1/runs/${run_id}/events`)
+            if (apiKey) eventsUrl.searchParams.set('token', apiKey)
+            logger.debug(`[AgentClients] ${this.name}: streaming events from ${eventsUrl}`)
+            const source = new EventSource(eventsUrl.toString())
+
+            let fullContent = ''
+
+            source.onmessage = (e: any) => {
+                try {
+                    const parsed = JSON.parse(e.data)
+                    logger.debug(`[AgentClients] ${this.name}: event=${parsed.event}`)
+
+                    if (parsed.event === 'run.completed') {
+                        source.close()
+                        logger.debug(`[AgentClients] ${this.name}: run completed, content length=${fullContent.length}`)
+                        if (fullContent) {
+                            this.stopTyping(roomId)
+                            this.sendMessage(roomId, fullContent)
+                        }
+                        this.deleteSession(actualSessionId).catch(() => { })
+                        onStatus?.('ready')
+                        return
+                    }
+
+                    if (parsed.event === 'run.failed') {
+                        source.close()
+                        logger.error(`[AgentClients] ${this.name}: run failed`)
+                        this.stopTyping(roomId)
+                        this.deleteSession(actualSessionId).catch(() => { })
+                        onStatus?.('ready')
+                        return
+                    }
+
+                    // Accumulate message deltas
+                    if (parsed.event === 'message.delta' && parsed.delta) {
+                        fullContent += parsed.delta
+                    }
+                } catch {
+                    // ignore parse errors
+                }
+            }
+
+            source.onerror = (err: any) => {
+                logger.error(err, `[AgentClients] ${this.name}: EventSource error`)
+                source.close()
+                this.stopTyping(roomId)
+                this.deleteSession(actualSessionId).catch(() => { })
+                onStatus?.('ready')
+            }
+        } catch (err: any) {
+            logger.error(`[AgentClients] ${this.name}: error handling message: ${err.message}`)
+            this.stopTyping(roomId)
+            this.deleteSession(sessionId).catch(() => { })
+            onStatus?.('ready')
+        }
+    }
+
+    private bindEvents(): void {
+        const s = this.socket!
+
+        s.on('typing', (data: any) => {
+            this.handlers.onTyping?.(data)
+        })
+
+        s.on('stop_typing', (data: any) => {
+            this.handlers.onStopTyping?.(data)
+        })
+
+        s.on('member_joined', (data: any) => {
+            this.handlers.onMemberJoined?.(data)
+        })
+
+        s.on('member_left', (data: any) => {
+            this.handlers.onMemberLeft?.(data)
+        })
+
+        // Auto rejoin rooms on reconnect
+        s.io.on('reconnect', async () => {
+            if (this._reconnecting) return
+            this._reconnecting = true
+            logger.info(`[AgentClients] ${this.name} reconnecting, rejoining ${this.joinedRooms.size} rooms...`)
+            const rooms = Array.from(this.joinedRooms)
+            for (const roomId of rooms) {
+                try {
+                    await this.joinRoom(roomId)
+                } catch (err: any) {
+                    logger.error(`[AgentClients] ${this.name} failed to rejoin room ${roomId}: ${err.message}`)
+                }
+            }
+            this._reconnecting = false
+        })
+    }
+}
+
+// ─── AgentClients (roomId -> agents) ──────────────────────────
+
+export class AgentClients {
+    private rooms = new Map<string, Map<string, AgentClient>>()
+    private _gatewayManager: GatewayManager | null = null
+    private _contextEngine: any = null
+    private _storage: any = null
+
+    // Per-room processing lock + mention queue
+    private _processingRooms = new Set<string>()
+    private _mentionQueue = new Map<string, Array<{ agent: AgentClient; msg: { content: string; senderName: string; senderId: string; timestamp: number } }>>()
+
+    /**
+     * Create an agent client and connect it to the server.
+     * The agent will NOT auto-join any room — call addAgentToRoom separately.
+     */
+    async createAgent(config: AgentConfig, handlers?: AgentEventHandler, port?: number): Promise<AgentClient> {
+        const client = new AgentClient(config, handlers)
+        await client.connect(port)
+
+        // Auto-apply stored references (fixes propagation for agents created after set*)
+        if (this._gatewayManager) client.setGatewayManager(this._gatewayManager)
+        if (this._contextEngine) client.setContextEngine(this._contextEngine)
+        if (this._storage) client.setStorage(this._storage)
+
+        logger.info(`[AgentClients] Connected: ${client.name} (${client.agentId})`)
+        return client
+    }
+
+    /**
+     * Connect an agent to a room.
+     */
+    async addAgentToRoom(roomId: string, client: AgentClient): Promise<JoinResult> {
+        let room = this.rooms.get(roomId)
+        if (!room) {
+            room = new Map()
+            this.rooms.set(roomId, room)
+        }
+
+        room.set(client.agentId, client)
+        const result = await client.joinRoom(roomId)
+        logger.info(`[AgentClients] ${client.name} joined room: ${roomId}`)
+        return result
+    }
+
+    /**
+     * Remove an agent from a room and disconnect it.
+     */
+    removeAgentFromRoom(roomId: string, agentId: string): void {
+        const room = this.rooms.get(roomId)
+        if (!room) return
+
+        const client = room.get(agentId)
+        if (client) {
+            client.disconnect()
+            room.delete(agentId)
+            logger.info(`[AgentClients] ${client.name} left room: ${roomId}`)
+
+            // Invalidate context engine cache for this agent
+            if (this._contextEngine) {
+                try { this._contextEngine.invalidateRoom(roomId) } catch { /* ignore */ }
+            }
+        }
+
+        if (room.size === 0) {
+            this.rooms.delete(roomId)
+        }
+    }
+
+    /**
+     * Get all agents in a room.
+     */
+    getAgents(roomId: string): AgentClient[] {
+        const room = this.rooms.get(roomId)
+        return room ? Array.from(room.values()) : []
+    }
+
+    /**
+     * Get a specific agent in a room.
+     */
+    getAgent(roomId: string, agentId: string): AgentClient | undefined {
+        return this.rooms.get(roomId)?.get(agentId)
+    }
+
+    /**
+     * Get all room IDs that have agents.
+     */
+    getRoomIds(): string[] {
+        return Array.from(this.rooms.keys())
+    }
+
+    /**
+     * Send a message from a specific agent in a room.
+     */
+    async sendMessage(roomId: string, agentId: string, content: string): Promise<string> {
+        const client = this.getAgent(roomId, agentId)
+        if (!client) {
+            throw new Error(`Agent "${agentId}" not found in room "${roomId}"`)
+        }
+        return client.sendMessage(roomId, content)
+    }
+
+    /**
+     * Broadcast a message from all agents in a room.
+     */
+    async broadcastFromRoom(roomId: string, content: string): Promise<string[]> {
+        const agents = this.getAgents(roomId)
+        return Promise.all(agents.map((agent) => agent.sendMessage(roomId, content)))
+    }
+
+    /**
+     * Disconnect all agents in a room.
+     */
+    disconnectRoom(roomId: string): void {
+        const room = this.rooms.get(roomId)
+        if (!room) return
+
+        room.forEach((client) => client.disconnect())
+        this.rooms.delete(roomId)
+        logger.info(`[AgentClients] All agents disconnected from room: ${roomId}`)
+
+        // Invalidate context engine cache for this room
+        if (this._contextEngine) {
+            try { this._contextEngine.invalidateRoom(roomId) } catch { /* ignore */ }
+        }
+    }
+
+    /**
+     * Disconnect all agents in all rooms.
+     */
+    disconnectAll(): void {
+        this.rooms.forEach((room) => {
+            room.forEach((client) => client.disconnect())
+        })
+        this.rooms.clear()
+        logger.info('[AgentClients] All agents disconnected')
+    }
+
+    /**
+     * Set gateway manager for all existing and future agents.
+     */
+    setGatewayManager(manager: GatewayManager): void {
+        this._gatewayManager = manager
+        this.rooms.forEach((room) => {
+            room.forEach((client) => client.setGatewayManager(manager))
+        })
+    }
+
+    /**
+     * Set context engine for all existing and future agents.
+     */
+    setContextEngine(engine: any): void {
+        this._contextEngine = engine
+        this.rooms.forEach((room) => {
+            room.forEach((client) => client.setContextEngine(engine))
+        })
+    }
+
+    /**
+     * Set message storage for all existing and future agents.
+     */
+    setStorage(storage: any): void {
+        this._storage = storage
+        this.rooms.forEach((room) => {
+            room.forEach((client) => client.setStorage(storage))
+        })
+    }
+
+
+    /**
+     * Server-side: parse @mentions and forward to matching agents directly.
+     * If the room is already processing (compressing/replying), queue the mention.
+     */
+    async processMentions(roomId: string, msg: { content: string; senderName: string; senderId: string; timestamp: number }): Promise<void> {
+        if (!this._gatewayManager) return
+
+        const content = msg.content.toLowerCase()
+        const agents = this.getAgents(roomId)
+
+        const mentioned = agents.filter(a => content.includes(`@${a.name.toLowerCase()}`))
+        if (mentioned.length === 0) return
+
+        logger.debug(`[AgentClients] ${mentioned.map(a => a.name).join(', ')} mentioned by ${msg.senderName}`)
+
+        for (const agent of mentioned) {
+            this._processAgentMention(roomId, agent, msg).catch((err) => {
+                logger.error(`[AgentClients] error processing mention for ${agent.name}: ${err.message}`)
+            })
+        }
+    }
+
+    /**
+     * Process a single agent mention with status reporting and queue drain.
+     */
+    private async _processAgentMention(
+        roomId: string,
+        agent: AgentClient,
+        msg: { content: string; senderName: string; senderId: string; timestamp: number },
+    ): Promise<void> {
+        const agentKey = `${roomId}:${agent.name}`
+        if (this._processingRooms.has(agentKey)) {
+            // Queue for this specific agent
+            let queue = this._mentionQueue.get(agentKey)
+            if (!queue) {
+                queue = []
+                this._mentionQueue.set(agentKey, queue)
+            }
+            queue.push({ agent, msg })
+            logger.debug(`[AgentClients] agent ${agent.name} is processing, queued mention in room ${roomId}`)
+            return
+        }
+
+        this._processingRooms.add(agentKey)
+        const onStatus = (status: 'compressing' | 'replying' | 'ready') => {
+            agent.emitContextStatus(roomId, status)
+            logger.debug(`[AgentClients] room ${roomId} agent ${agent.name} status: ${status}`)
+        }
+
+        try {
+            await agent.replyToMention(roomId, msg, onStatus)
+        } finally {
+            this._processingRooms.delete(agentKey)
+            await this._drainQueue(agentKey, roomId)
+        }
+    }
+
+    /**
+     * Drain queued mentions for a room after processing completes.
+     */
+    private async _drainQueue(agentKey: string, roomId: string): Promise<void> {
+        const queue = this._mentionQueue.get(agentKey)
+        if (!queue || queue.length === 0) return
+
+        this._mentionQueue.delete(agentKey)
+        logger.debug(`[AgentClients] draining ${queue.length} queued mention(s) for ${agentKey}`)
+
+        // Process the last queued mention only (most recent, discards stale intermediate ones)
+        const last = queue[queue.length - 1]
+        this._processingRooms.add(agentKey)
+        this._processAgentMention(roomId, last.agent, last.msg).catch((err) => {
+            logger.error(`[AgentClients] error processing queued mention: ${err.message}`)
+        })
+    }
+}
diff --git a/packages/server/src/services/hermes/group-chat/index.ts b/packages/server/src/services/hermes/group-chat/index.ts
new file mode 100644
index 0000000..a94400a
--- /dev/null
+++ b/packages/server/src/services/hermes/group-chat/index.ts
@@ -0,0 +1,866 @@
+import { Server, Socket, Namespace } from 'socket.io'
+import type { Server as HttpServer } from 'http'
+import { getToken } from '../../../services/auth'
+import { logger } from '../../../services/logger'
+import { getDb, ensureTable } from '../../../db'
+import { AgentClients } from './agent-clients'
+import { deleteSession as hermesDeleteSession } from '../hermes-cli'
+import { ContextEngine } from '../context-engine/compressor'
+
+// ─── Types ────────────────────────────────────────────────────
+
+interface ChatMessage {
+    id: string
+    roomId: string
+    senderId: string
+    senderName: string
+    content: string
+    timestamp: number
+}
+
+interface RoomAgent {
+    id: string
+    roomId: string
+    agentId: string
+    profile: string
+    name: string
+    description: string
+    invited: number
+}
+
+interface Member {
+    id: string
+    userId: string
+    name: string
+    description: string
+    joinedAt: number
+    online: boolean
+    socketId: string
+}
+
+// ─── SQLite Storage (global DB) ──────────────────────────────
+
+const GC_PENDING_SESSION_DELETES_SCHEMA: Record<string, string> = {
+    session_id: 'TEXT PRIMARY KEY',
+    profile_name: 'TEXT NOT NULL',
+    status: "TEXT NOT NULL DEFAULT 'pending'",
+    attempt_count: 'INTEGER NOT NULL DEFAULT 0',
+    last_error: 'TEXT',
+    created_at: 'INTEGER NOT NULL',
+    updated_at: 'INTEGER NOT NULL',
+    next_attempt_at: 'INTEGER NOT NULL DEFAULT 0',
+}
+
+const GC_SESSION_PROFILES_SCHEMA: Record<string, string> = {
+    session_id: 'TEXT PRIMARY KEY',
+    room_id: 'TEXT NOT NULL',
+    agent_id: 'TEXT NOT NULL',
+    profile_name: 'TEXT NOT NULL',
+    created_at: 'INTEGER NOT NULL',
+}
+
+const GC_ROOMS_SCHEMA: Record<string, string> = {
+    id: 'TEXT PRIMARY KEY',
+    name: 'TEXT NOT NULL',
+    inviteCode: 'TEXT UNIQUE',
+    triggerTokens: 'INTEGER NOT NULL DEFAULT 100000',
+    maxHistoryTokens: 'INTEGER NOT NULL DEFAULT 32000',
+    tailMessageCount: 'INTEGER NOT NULL DEFAULT 20',
+    totalTokens: 'INTEGER NOT NULL DEFAULT 0',
+}
+
+const GC_MESSAGES_SCHEMA: Record<string, string> = {
+    id: 'TEXT PRIMARY KEY',
+    roomId: 'TEXT NOT NULL',
+    senderId: 'TEXT NOT NULL',
+    senderName: 'TEXT NOT NULL',
+    content: 'TEXT NOT NULL',
+    timestamp: 'INTEGER NOT NULL',
+}
+
+const GC_ROOM_AGENTS_SCHEMA: Record<string, string> = {
+    id: 'TEXT PRIMARY KEY',
+    roomId: 'TEXT NOT NULL',
+    agentId: 'TEXT NOT NULL',
+    profile: 'TEXT NOT NULL',
+    name: 'TEXT NOT NULL',
+    description: "TEXT NOT NULL DEFAULT ''",
+    invited: 'INTEGER NOT NULL DEFAULT 0',
+}
+
+const GC_CONTEXT_SNAPSHOTS_SCHEMA: Record<string, string> = {
+    roomId: 'TEXT PRIMARY KEY',
+    summary: 'TEXT NOT NULL DEFAULT \'\'',
+    lastMessageId: 'TEXT NOT NULL',
+    lastMessageTimestamp: 'INTEGER NOT NULL',
+    updatedAt: 'INTEGER NOT NULL',
+}
+
+const GC_ROOM_MEMBERS_SCHEMA: Record<string, string> = {
+    id: 'TEXT PRIMARY KEY',
+    roomId: 'TEXT NOT NULL',
+    userId: 'TEXT NOT NULL',
+    userName: 'TEXT NOT NULL',
+    description: "TEXT NOT NULL DEFAULT ''",
+    joinedAt: 'INTEGER NOT NULL',
+    updatedAt: 'INTEGER NOT NULL',
+}
+
+let _tablesEnsured = false
+
+interface PendingSessionDelete {
+    session_id: string
+    profile_name: string
+    status: string
+    attempt_count: number
+    last_error: string | null
+    created_at: number
+    updated_at: number
+    next_attempt_at: number
+}
+
+interface GroupChatSessionProfile {
+    session_id: string
+    room_id: string
+    agent_id: string
+    profile_name: string
+    created_at: number
+}
+
+export interface PendingSessionDeleteDrainResult {
+    deleted: string[]
+    failed: Array<{ sessionId: string; error: string }>
+}
+
+class ChatStorage {
+    private db() { return getDb() }
+
+    init(): void {
+        if (_tablesEnsured) return
+        const db = this.db()
+        if (!db) return
+        ensureTable('gc_rooms', GC_ROOMS_SCHEMA)
+        ensureTable('gc_messages', GC_MESSAGES_SCHEMA)
+        ensureTable('gc_room_agents', GC_ROOM_AGENTS_SCHEMA)
+        ensureTable('gc_context_snapshots', GC_CONTEXT_SNAPSHOTS_SCHEMA)
+        ensureTable('gc_room_members', GC_ROOM_MEMBERS_SCHEMA)
+        ensureTable('gc_pending_session_deletes', GC_PENDING_SESSION_DELETES_SCHEMA)
+        ensureTable('gc_session_profiles', GC_SESSION_PROFILES_SCHEMA)
+        // Indexes (safe to run multiple times — CREATE INDEX IF NOT EXISTS)
+        try { db.exec('CREATE INDEX IF NOT EXISTS idx_gc_messages_room ON gc_messages(roomId, timestamp)') } catch { /* ignore */ }
+        try { db.exec('CREATE INDEX IF NOT EXISTS idx_gc_room_agents_room ON gc_room_agents(roomId)') } catch { /* ignore */ }
+        try { db.exec('CREATE UNIQUE INDEX IF NOT EXISTS idx_gc_room_members_unique ON gc_room_members(roomId, userId)') } catch { /* ignore */ }
+        try { db.exec('CREATE INDEX IF NOT EXISTS idx_gc_pending_session_deletes_profile ON gc_pending_session_deletes(profile_name, status, next_attempt_at, created_at)') } catch { /* ignore */ }
+        try { db.exec('CREATE INDEX IF NOT EXISTS idx_gc_session_profiles_profile ON gc_session_profiles(profile_name, created_at)') } catch { /* ignore */ }
+        _tablesEnsured = true
+    }
+
+    saveSessionProfile(sessionId: string, roomId: string, agentId: string, profileName: string): void {
+        this.db()?.prepare(
+            'INSERT INTO gc_session_profiles (session_id, room_id, agent_id, profile_name, created_at) VALUES (?, ?, ?, ?, ?) ON CONFLICT(session_id) DO UPDATE SET room_id = excluded.room_id, agent_id = excluded.agent_id, profile_name = excluded.profile_name'
+        ).run(sessionId, roomId, agentId, profileName, Date.now())
+    }
+
+    getSessionProfile(sessionId: string): GroupChatSessionProfile | null {
+        return (this.db()?.prepare(
+            'SELECT session_id, room_id, agent_id, profile_name, created_at FROM gc_session_profiles WHERE session_id = ?'
+        ).get(sessionId) as GroupChatSessionProfile | undefined) ?? null
+    }
+
+    deleteSessionProfile(sessionId: string): void {
+        this.db()?.prepare('DELETE FROM gc_session_profiles WHERE session_id = ?').run(sessionId)
+    }
+
+    listPendingSessionDeletes(profileName: string, limit = 50): PendingSessionDelete[] {
+        const rows = this.db()?.prepare(
+            `SELECT session_id, profile_name, status, attempt_count, last_error, created_at, updated_at, next_attempt_at
+             FROM gc_pending_session_deletes
+             WHERE profile_name = ? AND status = 'pending' AND next_attempt_at <= ?
+             ORDER BY created_at ASC
+             LIMIT ?`
+        ).all(profileName, Date.now(), limit) || []
+        return rows.map((row: any) => ({
+            session_id: String(row.session_id || ''),
+            profile_name: String(row.profile_name || ''),
+            status: String(row.status || 'pending'),
+            attempt_count: Number(row.attempt_count || 0),
+            last_error: row.last_error == null ? null : String(row.last_error),
+            created_at: Number(row.created_at || 0),
+            updated_at: Number(row.updated_at || 0),
+            next_attempt_at: Number(row.next_attempt_at || 0),
+        }))
+    }
+
+    enqueuePendingSessionDelete(sessionId: string, profileName: string): void {
+        const now = Date.now()
+        this.db()?.prepare(
+            `INSERT INTO gc_pending_session_deletes (session_id, profile_name, status, attempt_count, last_error, created_at, updated_at, next_attempt_at)
+             VALUES (?, ?, 'pending', 0, NULL, ?, ?, 0)
+             ON CONFLICT(session_id) DO UPDATE SET
+               profile_name = excluded.profile_name,
+               status = 'pending',
+               updated_at = excluded.updated_at,
+               next_attempt_at = 0`
+        ).run(sessionId, profileName, now, now)
+    }
+
+    claimPendingSessionDeletes(profileName: string, limit = 50): PendingSessionDelete[] {
+        const rows = this.listPendingSessionDeletes(profileName, limit)
+        if (rows.length === 0) return []
+        const now = Date.now()
+        const stmt = this.db()?.prepare(
+            `UPDATE gc_pending_session_deletes
+             SET status = 'processing', updated_at = ?
+             WHERE session_id = ? AND status = 'pending'`
+        )
+        const claimed: PendingSessionDelete[] = []
+        for (const row of rows) {
+            const result = stmt?.run(now, row.session_id)
+            if (result?.changes) {
+                claimed.push({ ...row, status: 'processing', updated_at: now })
+            }
+        }
+        return claimed
+    }
+
+    markPendingSessionDeleteFailed(sessionId: string, error: string): void {
+        const now = Date.now()
+        this.db()?.prepare(
+            `UPDATE gc_pending_session_deletes
+             SET status = 'pending',
+                 attempt_count = attempt_count + 1,
+                 last_error = ?,
+                 updated_at = ?,
+                 next_attempt_at = ?
+             WHERE session_id = ?`
+        ).run(error, now, now + 60_000, sessionId)
+    }
+
+    removePendingSessionDelete(sessionId: string): void {
+        this.db()?.prepare('DELETE FROM gc_pending_session_deletes WHERE session_id = ?').run(sessionId)
+    }
+
+    getPendingDeletedSessionIds(): Set<string> {
+        const rows = (this.db()?.prepare(
+            `SELECT session_id FROM gc_pending_session_deletes WHERE status IN ('pending', 'processing')`
+        ).all() || []) as Array<{ session_id: string }>
+        return new Set(rows.map(row => row.session_id))
+    }
+
+    // ─── Rooms ────────────────────────────────────────────────
+
+    getRoom(roomId: string): { id: string; name: string; inviteCode: string | null; triggerTokens: number; maxHistoryTokens: number; tailMessageCount: number; totalTokens: number } | undefined {
+        return this.db()?.prepare('SELECT id, name, inviteCode, triggerTokens, maxHistoryTokens, tailMessageCount, totalTokens FROM gc_rooms WHERE id = ?').get(roomId) as any
+    }
+
+    getRoomByInviteCode(code: string): { id: string; name: string; inviteCode: string | null; triggerTokens: number; maxHistoryTokens: number; tailMessageCount: number; totalTokens: number } | undefined {
+        return this.db()?.prepare('SELECT id, name, inviteCode, triggerTokens, maxHistoryTokens, tailMessageCount, totalTokens FROM gc_rooms WHERE inviteCode = ?').get(code) as any
+    }
+
+    getAllRooms(): { id: string; name: string; inviteCode: string | null; triggerTokens: number; maxHistoryTokens: number; tailMessageCount: number; totalTokens: number }[] {
+        return (this.db()?.prepare('SELECT id, name, inviteCode, triggerTokens, maxHistoryTokens, tailMessageCount, totalTokens FROM gc_rooms ORDER BY id').all() || []) as any[]
+    }
+
+    saveRoom(id: string, name: string, inviteCode?: string, config?: { triggerTokens?: number; maxHistoryTokens?: number; tailMessageCount?: number }): void {
+        this.db()?.prepare(
+            'INSERT OR IGNORE INTO gc_rooms (id, name, inviteCode, triggerTokens, maxHistoryTokens, tailMessageCount) VALUES (?, ?, ?, ?, ?, ?)'
+        ).run(id, name, inviteCode || null, config?.triggerTokens ?? 100000, config?.maxHistoryTokens ?? 32000, config?.tailMessageCount ?? 20)
+    }
+
+    updateRoomConfig(roomId: string, config: { triggerTokens?: number; maxHistoryTokens?: number; tailMessageCount?: number }): void {
+        const sets: string[] = []
+        const vals: any[] = []
+        if (config.triggerTokens !== undefined) { sets.push('triggerTokens = ?'); vals.push(config.triggerTokens) }
+        if (config.maxHistoryTokens !== undefined) { sets.push('maxHistoryTokens = ?'); vals.push(config.maxHistoryTokens) }
+        if (config.tailMessageCount !== undefined) { sets.push('tailMessageCount = ?'); vals.push(config.tailMessageCount) }
+        if (sets.length === 0) return
+        vals.push(roomId)
+        this.db()?.prepare(`UPDATE gc_rooms SET ${sets.join(', ')} WHERE id = ?`).run(...vals)
+    }
+
+    updateRoomInviteCode(roomId: string, inviteCode: string): void {
+        this.db()?.prepare('UPDATE gc_rooms SET inviteCode = ? WHERE id = ?').run(inviteCode, roomId)
+    }
+
+    updateRoomTotalTokens(roomId: string, tokens: number): void {
+        this.db()?.prepare('UPDATE gc_rooms SET totalTokens = ? WHERE id = ?').run(tokens, roomId)
+    }
+
+    estimateTokens(text: string): number {
+        const cjk = (text.match(/[\u2e80-\u9fff\uac00-\ud7af\u3000-\u303f\uff00-\uffef]/g) || []).length
+        const other = text.length - cjk
+        return Math.ceil(cjk * 1.5 + other / 4)
+    }
+
+    // ─── Messages ─────────────────────────────────────────────
+
+    getMessages(roomId: string, limit = 500): ChatMessage[] {
+        const rows = (this.db()?.prepare(
+            'SELECT id, roomId, senderId, senderName, content, timestamp FROM gc_messages WHERE roomId = ? ORDER BY timestamp DESC LIMIT ?'
+        ).all(roomId, limit) || []) as any[]
+        return rows.reverse()
+    }
+
+    addMessage(msg: ChatMessage): void {
+        this.db()?.prepare(
+            'INSERT INTO gc_messages (id, roomId, senderId, senderName, content, timestamp) VALUES (?, ?, ?, ?, ?, ?)'
+        ).run(msg.id, msg.roomId, msg.senderId, msg.senderName, msg.content, msg.timestamp)
+    }
+
+    pruneMessages(roomId: string, keep = 500): void {
+        const db = this.db()
+        if (!db) return
+        const count = (db.prepare('SELECT COUNT(*) as c FROM gc_messages WHERE roomId = ?').get(roomId) as any)?.c
+        if (count > keep) {
+            const cutoff = db.prepare(
+                'SELECT timestamp FROM gc_messages WHERE roomId = ? ORDER BY timestamp DESC LIMIT 1 OFFSET ?'
+            ).get(roomId, keep - 1) as any
+            if (cutoff) {
+                const result = db.prepare('DELETE FROM gc_messages WHERE roomId = ? AND timestamp < ?').run(roomId, cutoff.timestamp)
+                logger.info(`[GroupChat] pruned ${result.changes} messages from room ${roomId} (had ${count}, keeping ${keep})`)
+            }
+        }
+    }
+
+    // ─── Room Agents ──────────────────────────────────────────
+
+    getRoomAgents(roomId: string): RoomAgent[] {
+        return (this.db()?.prepare(
+            'SELECT id, roomId, agentId, profile, name, description, invited FROM gc_room_agents WHERE roomId = ?'
+        ).all(roomId) || []) as unknown as RoomAgent[]
+    }
+
+    addRoomAgent(roomId: string, agentId: string, profile: string, name: string, description: string, invited: number): RoomAgent {
+        const id = Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
+        this.db()?.prepare(
+            'INSERT INTO gc_room_agents (id, roomId, agentId, profile, name, description, invited) VALUES (?, ?, ?, ?, ?, ?, ?)'
+        ).run(id, roomId, agentId, profile, name, description, invited)
+        return { id, roomId, agentId, profile, name, description, invited }
+    }
+
+    removeRoomAgent(agentId: string): void {
+        this.db()?.prepare('DELETE FROM gc_room_agents WHERE id = ?').run(agentId)
+    }
+
+    // ─── Context Snapshots ──────────────────────────────────
+
+    getContextSnapshot(roomId: string): { roomId: string; summary: string; lastMessageId: string; lastMessageTimestamp: number; updatedAt: number } | null {
+        return (this.db()?.prepare(
+            'SELECT roomId, summary, lastMessageId, lastMessageTimestamp, updatedAt FROM gc_context_snapshots WHERE roomId = ?'
+        ).get(roomId) as any) ?? null
+    }
+
+    saveContextSnapshot(roomId: string, summary: string, lastMessageId: string, lastMessageTimestamp: number): void {
+        this.db()?.prepare(
+            'INSERT INTO gc_context_snapshots (roomId, summary, lastMessageId, lastMessageTimestamp, updatedAt) VALUES (?, ?, ?, ?, ?) ON CONFLICT(roomId) DO UPDATE SET summary = excluded.summary, lastMessageId = excluded.lastMessageId, lastMessageTimestamp = excluded.lastMessageTimestamp, updatedAt = excluded.updatedAt'
+        ).run(roomId, summary, lastMessageId, lastMessageTimestamp, Date.now())
+    }
+
+    deleteContextSnapshot(roomId: string): void {
+        this.db()?.prepare('DELETE FROM gc_context_snapshots WHERE roomId = ?').run(roomId)
+    }
+
+    deleteRoom(roomId: string): void {
+        const db = this.db()
+        if (!db) return
+        db.prepare('DELETE FROM gc_messages WHERE roomId = ?').run(roomId)
+        db.prepare('DELETE FROM gc_room_agents WHERE roomId = ?').run(roomId)
+        db.prepare('DELETE FROM gc_room_members WHERE roomId = ?').run(roomId)
+        db.prepare('DELETE FROM gc_context_snapshots WHERE roomId = ?').run(roomId)
+        db.prepare('DELETE FROM gc_rooms WHERE id = ?').run(roomId)
+    }
+
+    // ─── Room Members ──────────────────────────────────────
+
+    getRoomMembers(roomId: string): { id: string; userId: string; name: string; description: string; joinedAt: number }[] {
+        return (this.db()?.prepare(
+            'SELECT id, userId, userName as name, description, joinedAt FROM gc_room_members WHERE roomId = ? ORDER BY joinedAt'
+        ).all(roomId) || []) as unknown as { id: string; userId: string; name: string; description: string; joinedAt: number }[]
+    }
+
+    addRoomMember(roomId: string, userId: string, userName: string, description: string): void {
+        const existing = this.getMemberByUserId(roomId, userId)
+        if (existing) {
+            // Update name/description on rejoin, refresh updatedAt
+            this.db()?.prepare(
+                'UPDATE gc_room_members SET userName = ?, description = ?, updatedAt = ? WHERE roomId = ? AND userId = ?'
+            ).run(userName, description, Date.now(), roomId, userId)
+            return
+        }
+        const id = Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
+        const now = Date.now()
+        this.db()?.prepare(
+            'INSERT INTO gc_room_members (id, roomId, userId, userName, description, joinedAt, updatedAt) VALUES (?, ?, ?, ?, ?, ?, ?)'
+        ).run(id, roomId, userId, userName, description, now, now)
+    }
+
+    getMemberByUserId(roomId: string, userId: string): Member | null {
+        return (this.db()?.prepare(
+            'SELECT id, userId, userName as name, description, joinedAt FROM gc_room_members WHERE roomId = ? AND userId = ?'
+        ).get(roomId, userId) as any) ?? null
+    }
+
+    updateMemberActivity(roomId: string, userId: string): void {
+        this.db()?.prepare(
+            'UPDATE gc_room_members SET updatedAt = ? WHERE roomId = ? AND userId = ?'
+        ).run(Date.now(), roomId, userId)
+    }
+}
+
+export async function drainPendingSessionDeletes(profileName: string): Promise<PendingSessionDeleteDrainResult> {
+    const storage = new ChatStorage()
+    storage.init()
+    const claimed = storage.claimPendingSessionDeletes(profileName)
+    const result: PendingSessionDeleteDrainResult = { deleted: [], failed: [] }
+
+    for (const item of claimed) {
+        try {
+            const ok = await hermesDeleteSession(item.session_id)
+            if (!ok) {
+                throw new Error('Failed to delete session')
+            }
+            storage.removePendingSessionDelete(item.session_id)
+            storage.deleteSessionProfile(item.session_id)
+            result.deleted.push(item.session_id)
+        } catch (err: any) {
+            const message = err?.message || 'Failed to delete session'
+            storage.markPendingSessionDeleteFailed(item.session_id, message)
+            result.failed.push({ sessionId: item.session_id, error: message })
+        }
+    }
+
+    return result
+}
+
+// ─── ChatRoom (in-memory, for online members) ─────────────────
+
+class ChatRoom {
+    readonly id: string
+    name: string
+    readonly members = new Map<string, Member>()
+
+    constructor(id: string, name?: string) {
+        this.id = id
+        this.name = name || id
+    }
+
+    addOrUpdateMember(socketId: string, userId: string, name: string, description: string): Member {
+        const existing = this.members.get(userId)
+        if (existing) {
+            existing.name = name
+            existing.description = description
+            existing.online = true
+            existing.socketId = socketId
+            return existing
+        }
+        const member: Member = { id: socketId, userId, name, description, joinedAt: Date.now(), online: true, socketId }
+        this.members.set(userId, member)
+        return member
+    }
+
+    removeMember(socketId: string): void {
+        for (const member of this.members.values()) {
+            if (member.socketId === socketId) {
+                member.online = false
+                break
+            }
+        }
+    }
+
+    getMembersList(): Member[] {
+        return Array.from(this.members.values())
+    }
+
+    getOnlineMemberBySocketId(socketId: string): Member | undefined {
+        for (const member of this.members.values()) {
+            if (member.socketId === socketId && member.online) return member
+        }
+        return undefined
+    }
+
+    hasOnlineMember(socketId: string): boolean {
+        return this.getOnlineMemberBySocketId(socketId) !== undefined
+    }
+}
+
+// ─── GroupChat Server ────────────────────────────────────────
+
+export class GroupChatServer {
+    private io: Server
+    private nsp: Namespace
+    private storage: ChatStorage
+    private rooms = new Map<string, ChatRoom>()
+    /** Map: socket.id → persistent userId */
+    private socketUserMap = new Map<string, string>()
+    /** Map: userId → { name, description } (from auth) */
+    private userInfoMap = new Map<string, { name: string; description: string }>()
+    readonly agentClients = new AgentClients()
+    private _contextEngine: ContextEngine | null = null
+    private _restoreScheduled = false
+    /** roomId -> (userId -> { userName, timer }) */
+    private typingState = new Map<string, Map<string, { userName: string; timer: ReturnType<typeof setTimeout> }>>()
+    /** roomId -> (agentName -> { agentName, status }) */
+    private contextStatusState = new Map<string, Map<string, { agentName: string; status: string }>>()
+
+    setGatewayManager(manager: any): void {
+        this.agentClients.setGatewayManager(manager)
+        if (this._contextEngine && manager) {
+            this._contextEngine.setUpstream(manager.getUpstream(''), manager.getApiKey(''))
+        }
+    }
+
+    constructor(httpServer: HttpServer) {
+        this.storage = new ChatStorage()
+        this.storage.init()
+
+        this.io = new Server(httpServer, {
+            cors: { origin: '*' }
+        })
+        this.nsp = this.io.of('/group-chat')
+        this.nsp.use(this.authMiddleware.bind(this))
+        this.nsp.on('connection', this.onConnection.bind(this))
+
+        // Restore persisted rooms into memory
+        this.storage.getAllRooms().forEach((row) => {
+            this.rooms.set(row.id, new ChatRoom(row.id, row.name))
+        })
+
+        logger.info('[GroupChat] Socket.IO ready at /group-chat')
+
+        // Initialize context engine for group chat compression
+        const contextEngine = new ContextEngine({
+            messageFetcher: this.storage,
+            sessionCleaner: async (sessionId: string) => {
+                try {
+                    await hermesDeleteSession(sessionId)
+                } catch (err: any) {
+                    logger.warn(`[GroupChat] failed to delete compression session ${sessionId}: ${err.message}`)
+                }
+            },
+        })
+        this.agentClients.setContextEngine(contextEngine)
+        this.agentClients.setStorage(this.storage)
+        this._contextEngine = contextEngine
+
+        // Restore agent connections — call restoreAgents() after server is listening
+        this._restoreScheduled = false
+    }
+
+    getIO(): Server {
+        return this.io
+    }
+
+    getStorage(): ChatStorage {
+        return this.storage
+    }
+
+    getContextEngine(): ContextEngine | null {
+        return this._contextEngine || null
+    }
+
+    getRoomIds(): string[] {
+        return Array.from(this.rooms.keys())
+    }
+
+    // ─── Restore Agents ─────────────────────────────────────────
+
+    /**
+     * Restore persisted agent connections. Safe to call multiple times;
+     * will only execute once.
+     */
+    async restoreWhenReady(): Promise<void> {
+        if (this._restoreScheduled) return
+        this._restoreScheduled = true
+        await this.restoreAgents()
+    }
+
+    private async restoreAgents(): Promise<void> {
+        const rooms = this.storage.getAllRooms()
+        let total = 0
+
+        for (const room of rooms) {
+            const agents = this.storage.getRoomAgents(room.id)
+            for (const agent of agents) {
+                try {
+                    const client = await this.agentClients.createAgent({
+                        profile: agent.profile,
+                        name: agent.name,
+                        description: agent.description,
+                        invited: agent.invited,
+                    })
+                    await this.agentClients.addAgentToRoom(room.id, client)
+                    total++
+                } catch (err: any) {
+                    logger.error(`[GroupChat] Failed to restore agent ${agent.name} in room ${room.id}: ${err.message}`)
+                }
+            }
+        }
+
+        if (total > 0) {
+            logger.info(`[GroupChat] Restored ${total} agent(s) across ${rooms.length} room(s)`)
+        }
+    }
+
+    // ─── Auth ───────────────────────────────────────────────────
+
+    private async authMiddleware(socket: Socket, next: (err?: Error) => void): Promise<void> {
+        const authToken = await getToken()
+        const token = socket.handshake.auth.token || socket.handshake.query.token || ''
+        if (authToken) {
+            if (token !== authToken) {
+                return next(new Error('Unauthorized'))
+            }
+        }
+        next()
+    }
+
+    // ─── Connection ─────────────────────────────────────────────
+
+    private onConnection(socket: Socket): void {
+        const auth = socket.handshake.auth as { userId?: string; name?: string; description?: string }
+        const userId = auth.userId || socket.id
+        const userName = auth.name || `User-${userId.slice(0, 6)}`
+        const description = auth.description || ''
+
+        this.socketUserMap.set(socket.id, userId)
+        this.userInfoMap.set(userId, { name: userName, description })
+
+        logger.debug(`[GroupChat] Connected: ${userName} (socket=${socket.id}, user=${userId})`)
+
+        socket.on('join', (data: { roomId?: string; name?: string }, ack?: (response?: unknown) => void) => this.handleJoin(socket, data, ack))
+        socket.on('message', (data: { roomId?: string; content: string }, ack?: (response?: unknown) => void) => this.handleMessage(socket, data, ack))
+        socket.on('typing', (data: { roomId?: string }) => this.handleTyping(socket, data))
+        socket.on('stop_typing', (data: { roomId?: string }) => this.handleStopTyping(socket, data))
+        socket.on('context_status', (data: { roomId?: string; agentName?: string; status?: string }) => this.handleContextStatus(socket, data))
+        socket.on('disconnect', () => this.handleDisconnect(socket))
+    }
+
+    // ─── Handlers ───────────────────────────────────────────────
+
+    private handleJoin(socket: Socket, data: { roomId?: string; name?: string; description?: string }, ack?: (res: any) => void): void {
+        const socketId = socket.id
+        const userId = this.socketUserMap.get(socketId) || socketId
+        const userInfo = this.userInfoMap.get(userId) || { name: `User-${userId.slice(0, 6)}`, description: '' }
+        const userName = data.name || userInfo.name
+        const description = data.description || userInfo.description
+
+        // Update stored user info
+        this.userInfoMap.set(userId, { name: userName, description })
+
+        const roomId = data.roomId || 'general'
+        let room = this.rooms.get(roomId)
+        if (!room) {
+            room = new ChatRoom(roomId)
+            this.rooms.set(roomId, room)
+            this.storage.saveRoom(roomId, roomId)
+        }
+
+        // Persist member to SQLite
+        this.storage.addRoomMember(roomId, userId, userName, description)
+
+        // Add to in-memory online members (keyed by userId)
+        room.addOrUpdateMember(socketId, userId, userName, description)
+        socket.join(roomId)
+
+        socket.to(roomId).emit('member_joined', {
+            roomId,
+            memberId: userId,
+            memberName: userName,
+            members: room.getMembersList(),
+        })
+
+        // Load history from SQLite
+        const messages = this.storage.getMessages(roomId)
+        const agents = this.storage.getRoomAgents(roomId)
+
+        ack?.({
+            roomId,
+            roomName: room.name,
+            members: room.getMembersList(),
+            messages,
+            agents,
+            rooms: this.getRoomIds(),
+            typingUsers: this.getTypingUsers(roomId),
+            contextStatuses: this.getContextStatuses(roomId),
+        })
+
+        logger.debug(`[GroupChat] ${userName} (user=${userId}) joined room: ${roomId}`)
+    }
+
+    private handleMessage(socket: Socket, data: { roomId?: string; content: string }, ack?: (res: any) => void): void {
+        const socketId = socket.id
+        const roomId = data.roomId || 'general'
+        const room = this.rooms.get(roomId)
+
+        if (!room || !room.hasOnlineMember(socketId)) {
+            ack?.({ error: 'Not in room' })
+            return
+        }
+
+        const member = room.getOnlineMemberBySocketId(socketId)
+        const userId = member?.userId || socketId
+        const userName = member?.name || `User-${socketId.slice(0, 6)}`
+
+        const msg: ChatMessage = {
+            id: this.generateId(),
+            roomId,
+            senderId: userId,
+            senderName: userName,
+            content: data.content,
+            timestamp: Date.now(),
+        }
+
+        this.storage.addMessage(msg)
+        this.storage.pruneMessages(roomId)
+
+        // Recalculate total tokens for the room
+        const messages = this.storage.getMessages(roomId)
+        const totalTokens = this.storage.estimateTokens(messages.map(m => m.content + m.senderName).join(''))
+        this.storage.updateRoomTotalTokens(roomId, totalTokens)
+
+        this.nsp.to(roomId).emit('message', msg)
+        this.nsp.to(roomId).emit('room_updated', { roomId, totalTokens })
+        ack?.({ id: msg.id })
+
+        // Server-side @mention routing — parse mentions and invoke agents directly
+        this.agentClients.processMentions(roomId, {
+            content: msg.content,
+            senderName: msg.senderName,
+            senderId: msg.senderId,
+            timestamp: msg.timestamp,
+        }).catch((err) => {
+            logger.error(`[GroupChat] processMentions error: ${err.message}`)
+        })
+    }
+
+    private handleTyping(socket: Socket, data: { roomId?: string }): void {
+        const roomId = data.roomId || 'general'
+        const userId = this.socketUserMap.get(socket.id) || socket.id
+        const userName = this.userInfoMap.get(userId)?.name || `User-${socket.id.slice(0, 6)}`
+
+        // Track typing state for rejoin recovery
+        let roomTyping = this.typingState.get(roomId)
+        if (!roomTyping) {
+            roomTyping = new Map()
+            this.typingState.set(roomId, roomTyping)
+        }
+        const existing = roomTyping.get(userId)
+        if (existing) clearTimeout(existing.timer)
+        roomTyping.set(userId, {
+            userName,
+            timer: setTimeout(() => {
+                roomTyping!.delete(userId)
+                if (roomTyping!.size === 0) this.typingState.delete(roomId)
+            }, 30000),
+        })
+
+        socket.to(roomId).emit('typing', {
+            roomId,
+            userId,
+            userName,
+        })
+    }
+
+    private handleStopTyping(socket: Socket, data: { roomId?: string }): void {
+        const roomId = data.roomId || 'general'
+        const userId = this.socketUserMap.get(socket.id) || socket.id
+
+        // Remove from typing state
+        const roomTyping = this.typingState.get(roomId)
+        if (roomTyping) {
+            const entry = roomTyping.get(userId)
+            if (entry) clearTimeout(entry.timer)
+            roomTyping.delete(userId)
+            if (roomTyping.size === 0) this.typingState.delete(roomId)
+        }
+
+        socket.to(roomId).emit('stop_typing', {
+            roomId,
+            userId,
+        })
+    }
+
+    private handleContextStatus(socket: Socket, data: { roomId?: string; agentName?: string; status?: string }): void {
+        const roomId = data.roomId || 'general'
+        const agentName = data.agentName || ''
+        const status = data.status || ''
+
+        if (!agentName) return
+
+        let roomStatuses = this.contextStatusState.get(roomId)
+        if (!roomStatuses) {
+            roomStatuses = new Map()
+            this.contextStatusState.set(roomId, roomStatuses)
+        }
+
+        if (status === 'ready') {
+            roomStatuses.delete(agentName)
+            if (roomStatuses.size === 0) this.contextStatusState.delete(roomId)
+        } else {
+            roomStatuses.set(agentName, { agentName, status })
+        }
+
+        // Relay to all other sockets in the room
+        socket.to(roomId).emit('context_status', {
+            roomId,
+            agentName,
+            status,
+        })
+    }
+
+    private handleDisconnect(socket: Socket): void {
+        const socketId = socket.id
+        const userId = this.socketUserMap.get(socketId)
+        const userName = userId ? this.userInfoMap.get(userId)?.name : undefined
+
+        logger.debug(`[GroupChat] Disconnected: ${userName || socketId} (socket=${socketId}, user=${userId || socketId})`)
+
+        // Clean up typing state for this socket
+        for (const [roomId, roomTyping] of this.typingState) {
+            const entry = roomTyping.get(userId || socketId)
+            if (entry) {
+                clearTimeout(entry.timer)
+                roomTyping.delete(userId || socketId)
+                if (roomTyping.size === 0) this.typingState.delete(roomId)
+            }
+        }
+
+        this.leaveAllRooms(socket, socketId)
+        this.socketUserMap.delete(socketId)
+        // Don't delete userInfoMap — it persists across reconnects
+    }
+
+    // ─── Helpers ────────────────────────────────────────────────
+
+    private getTypingUsers(roomId: string): Array<{ userId: string; userName: string }> {
+        const roomTyping = this.typingState.get(roomId)
+        if (!roomTyping) return []
+        return Array.from(roomTyping.entries()).map(([userId, entry]) => ({ userId, userName: entry.userName }))
+    }
+
+    private getContextStatuses(roomId: string): Array<{ agentName: string; status: string }> {
+        const roomStatuses = this.contextStatusState.get(roomId)
+        if (!roomStatuses) return []
+        return Array.from(roomStatuses.values())
+    }
+
+    private leaveAllRooms(socket: Socket, socketId: string): void {
+        this.rooms.forEach((room, rid) => {
+            if (room.hasOnlineMember(socketId)) {
+                const member = room.getOnlineMemberBySocketId(socketId)
+                room.removeMember(socketId)
+                socket.leave(rid)
+                this.nsp.to(rid).emit('member_left', {
+                    roomId: rid,
+                    memberId: member?.userId || socketId,
+                    memberName: member?.name || `User-${socketId.slice(0, 6)}`,
+                    members: room.getMembersList(),
+                })
+            }
+        })
+    }
+
+    private generateId(): string {
+        return Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
+    }
+}
diff --git a/packages/server/src/services/logger.ts b/packages/server/src/services/logger.ts
index fa68352..042b11a 100644
--- a/packages/server/src/services/logger.ts
+++ b/packages/server/src/services/logger.ts
@@ -23,7 +23,7 @@ function rotateIfNeeded() {
       truncateSync(logFile, 0)
       writeFileSync(logFile, buf)
     }
-  } catch {}
+  } catch { }
 }
 
 // Rotate on startup
@@ -36,5 +36,5 @@ export const logger = pino({
   level: process.env.LOG_LEVEL || 'info',
 }, pino.destination({
   dest: logFile,
-  sync: false,
+  sync: true,
 }))
diff --git a/packages/server/src/services/shutdown.ts b/packages/server/src/services/shutdown.ts
index 52b2baa..0c297ca 100644
--- a/packages/server/src/services/shutdown.ts
+++ b/packages/server/src/services/shutdown.ts
@@ -1,6 +1,6 @@
 import { logger } from './logger'
 
-export function bindShutdown(server: any): void {
+export function bindShutdown(server: any, groupChatServer?: any): void {
   let isShuttingDown = false
 
   const shutdown = async (signal: string) => {
@@ -10,6 +10,13 @@ export function bindShutdown(server: any): void {
     logger.info('Shutting down (%s)...', signal)
 
     try {
+      // Disconnect Socket.IO before HTTP server to prevent hanging
+      if (groupChatServer) {
+        groupChatServer.agentClients.disconnectAll()
+        groupChatServer.getIO().close()
+        logger.info('Socket.IO closed')
+      }
+
       if (server) {
         await new Promise<void>((resolve) => {
           server.close(() => {
diff --git a/scripts/build-server.mjs b/scripts/build-server.mjs
index b86daa3..c223d9d 100644
--- a/scripts/build-server.mjs
+++ b/scripts/build-server.mjs
@@ -14,12 +14,12 @@ await esbuild.build({
   target: 'node23',
   format: 'cjs',
   outfile: resolve(rootDir, 'dist/server/index.js'),
-  external: ['node-pty', 'node:sqlite'],
+  external: ['node-pty', 'node:sqlite', 'socket.io'],
   define: {
     __APP_VERSION__: JSON.stringify(version),
   },
   sourcemap: true,
-  minify: false,
-  treeShaking: false,
+  minify: true,
+  treeShaking: true,
   logLevel: 'info',
 })
diff --git a/tests/client/chat-store.test.ts b/tests/client/chat-store.test.ts
index 4a61787..cf1ccb4 100644
--- a/tests/client/chat-store.test.ts
+++ b/tests/client/chat-store.test.ts
@@ -94,6 +94,8 @@ describe('Chat Store', () => {
     window.localStorage.setItem(ACTIVE_SESSION_KEY, 'local-1')
     window.localStorage.setItem(SESSIONS_CACHE_KEY, JSON.stringify([cachedSession]))
     window.localStorage.setItem(sessionMessagesKey('local-1'), JSON.stringify(cachedMessages))
+    // Mark local-1 as in-flight so loadSessions preserves it
+    window.localStorage.setItem(inFlightKey('local-1'), JSON.stringify({ runId: 'run-1', startedAt: Date.now() }))
 
     mockSessionsApi.fetchSessions.mockResolvedValue([makeSummary('remote-1', 'Remote Session')])
     mockSessionsApi.fetchSession.mockResolvedValue(null)
diff --git a/tests/client/markdown-rendering.test.ts b/tests/client/markdown-rendering.test.ts
index 09f4e0b..cddd666 100644
--- a/tests/client/markdown-rendering.test.ts
+++ b/tests/client/markdown-rendering.test.ts
@@ -8,6 +8,15 @@ vi.mock('vue-i18n', () => ({
   }),
 }))
 
+vi.mock('naive-ui', () => ({
+  useMessage: () => ({
+    error: vi.fn(),
+    success: vi.fn(),
+    warning: vi.fn(),
+    info: vi.fn(),
+  }),
+}))
+
 import MarkdownRenderer from '@/components/hermes/chat/MarkdownRenderer.vue'
 
 describe('MarkdownRenderer', () => {
diff --git a/tests/client/message-item-highlight.test.ts b/tests/client/message-item-highlight.test.ts
index 42d6fe4..0e6cb8c 100644
--- a/tests/client/message-item-highlight.test.ts
+++ b/tests/client/message-item-highlight.test.ts
@@ -8,6 +8,15 @@ vi.mock('vue-i18n', () => ({
   }),
 }))
 
+vi.mock('naive-ui', () => ({
+  useMessage: () => ({
+    error: vi.fn(),
+    success: vi.fn(),
+    warning: vi.fn(),
+    info: vi.fn(),
+  }),
+}))
+
 import MessageItem from '@/components/hermes/chat/MessageItem.vue'
 import type { Message } from '@/stores/hermes/chat'
 
diff --git a/tests/server/context-engine.test.ts b/tests/server/context-engine.test.ts
new file mode 100644
index 0000000..8d7f17f
--- /dev/null
+++ b/tests/server/context-engine.test.ts
@@ -0,0 +1,455 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+import { SummaryCache } from '../../packages/server/src/services/hermes/context-engine/summary-cache'
+import {
+    buildAgentInstructions,
+    buildSummarizationSystemPrompt,
+    buildFullSummaryPrompt,
+    buildIncrementalUpdatePrompt,
+} from '../../packages/server/src/services/hermes/context-engine/prompt'
+import { ContextEngine } from '../../packages/server/src/services/hermes/context-engine/compressor'
+import type { StoredMessage, MessageFetcher, GatewayCaller } from '../../packages/server/src/services/hermes/context-engine/types'
+
+// ─── Helpers ─────────────────────────────────────────────────
+
+function makeMessage(overrides: Partial<StoredMessage> = {}): StoredMessage {
+    return {
+        id: 'msg-1',
+        roomId: 'room-1',
+        senderId: 'user-1',
+        senderName: 'Alice',
+        content: 'Hello world',
+        timestamp: 1000,
+        ...overrides,
+    }
+}
+
+function makeMessages(count: number, roomId = 'room-1', startTimestamp = 1000): StoredMessage[] {
+    return Array.from({ length: count }, (_, i) => makeMessage({
+        id: `msg-${i}`,
+        roomId,
+        senderId: i % 3 === 0 ? 'agent-socket' : `user-${i}`,
+        senderName: i % 3 === 0 ? 'Claude' : `User${i}`,
+        content: `Message ${i} with some content`,
+        timestamp: startTimestamp + i * 1000,
+    }))
+}
+
+// ─── SummaryCache ─────────────────────────────────────────────
+
+describe('SummaryCache', () => {
+    it('stores and retrieves entries', () => {
+        const cache = new SummaryCache(60_000)
+        cache.set('room-1', {
+            summary: 'Summary text',
+            lastMessageId: 'msg-10',
+            lastMessageTimestamp: 5000,
+            createdAt: Date.now(),
+        })
+        const entry = cache.get('room-1')
+        expect(entry).toBeDefined()
+        expect(entry!.summary).toBe('Summary text')
+    })
+
+    it('returns undefined for expired entries', () => {
+        const cache = new SummaryCache(100) // 100ms TTL
+        cache.set('room-1', {
+            summary: 'Old summary',
+            lastMessageId: 'msg-5',
+            lastMessageTimestamp: 5000,
+            createdAt: Date.now() - 200, // created 200ms ago
+        })
+        expect(cache.get('room-1')).toBeUndefined()
+    })
+
+    it('invalidates entries for a room', () => {
+        const cache = new SummaryCache(60_000)
+        cache.set('room-1', { summary: 'A', lastMessageId: 'msg-1', lastMessageTimestamp: 1000, createdAt: Date.now() })
+        cache.set('room-2', { summary: 'C', lastMessageId: 'msg-3', lastMessageTimestamp: 3000, createdAt: Date.now() })
+
+        cache.invalidate('room-1')
+        expect(cache.get('room-1')).toBeUndefined()
+        expect(cache.get('room-2')).toBeDefined()
+    })
+
+    it('enforces max entry limit', () => {
+        const cache = new SummaryCache(60_000)
+        // Fill cache beyond limit (internal MAX_ENTRIES = 200)
+        for (let i = 0; i < 210; i++) {
+            cache.set(`room-${i}`, {
+                summary: `Summary ${i}`,
+                lastMessageId: `msg-${i}`,
+                lastMessageTimestamp: i * 1000,
+                createdAt: Date.now() - (210 - i), // earlier entries have older createdAt
+            })
+        }
+        // Cache should not exceed 200 entries
+        expect(cache.size).toBeLessThanOrEqual(200)
+    })
+})
+
+// ─── Prompts ──────────────────────────────────────────────────
+
+describe('prompts', () => {
+    it('builds agent instructions with all fields', () => {
+        const result = buildAgentInstructions({
+            agentName: 'Claude',
+            roomName: 'general',
+            agentDescription: 'AI coding assistant',
+            memberNames: ['Alice', 'Bob', 'Claude'],
+            members: [
+                { userId: 'u1', name: 'Alice', description: 'dev' },
+                { userId: 'u2', name: 'Bob', description: 'designer' },
+                { userId: 'u3', name: 'Claude', description: '' },
+            ],
+        })
+        expect(result).toContain('"Claude"')
+        expect(result).toContain('general')
+        expect(result).toContain('AI coding assistant')
+        expect(result).toContain('Alice')
+        expect(result).toContain('Bob')
+        expect(result).toContain('@Claude')
+    })
+
+    it('builds agent instructions with empty member list', () => {
+        const result = buildAgentInstructions({
+            agentName: 'GPT',
+            roomName: 'dev',
+            agentDescription: 'Helper',
+            memberNames: [],
+            members: [],
+        })
+        expect(result).toContain('"GPT"')
+        expect(result).toContain('未知')
+    })
+
+    it('builds agent instructions using memberNames when members is empty', () => {
+        const result = buildAgentInstructions({
+            agentName: 'GPT',
+            roomName: 'dev',
+            agentDescription: 'Helper',
+            memberNames: ['Alice', 'Bob'],
+            members: [],
+        })
+        expect(result).toContain('Alice')
+        expect(result).toContain('Bob')
+    })
+
+    it('builds summarization system prompt', () => {
+        const result = buildSummarizationSystemPrompt()
+        expect(result).toContain('摘要')
+    })
+
+    it('builds full summary prompt', () => {
+        const result = buildFullSummaryPrompt()
+        expect(result).toContain('摘要')
+    })
+
+    it('builds incremental update prompt', () => {
+        const result = buildIncrementalUpdatePrompt()
+        expect(result).toContain('更新')
+    })
+})
+
+// ─── ContextEngine.buildContext ────────────────────────────────
+
+describe('ContextEngine.buildContext', () => {
+    let mockSummarize = vi.fn().mockResolvedValue({ summary: 'Summary of conversation.', sessionId: 'comp-1' })
+    const mockGatewayCaller: GatewayCaller = {
+        summarize: mockSummarize,
+    }
+
+    let mockFetcher: MessageFetcher
+    let engine: ContextEngine
+
+    beforeEach(() => {
+        vi.clearAllMocks()
+        mockFetcher = {
+            getMessages: vi.fn().mockReturnValue([]),
+            getContextSnapshot: vi.fn().mockReturnValue(null),
+            saveContextSnapshot: vi.fn(),
+            deleteContextSnapshot: vi.fn(),
+        }
+        engine = new ContextEngine({
+            config: { maxHistoryTokens: 4000, tailMessageCount: 10, triggerTokens: 100_000, charsPerToken: 4, summarizationTimeoutMs: 30_000 },
+            messageFetcher: mockFetcher,
+            gatewayCaller: { summarize: mockSummarize },
+        })
+    })
+
+    it('returns all messages as history when under threshold', async () => {
+        const messages = makeMessages(10) // 10 messages, under trigger threshold
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        const result = await engine.buildContext({
+            roomId: 'room-1',
+            agentId: 'agent-1',
+            agentName: 'Claude',
+            agentDescription: 'Helper',
+            agentSocketId: 'agent-socket',
+            roomName: 'general',
+            memberNames: ['Alice'],
+            members: [{ userId: 'u1', name: 'Alice', description: '' }],
+            upstream: 'http://localhost:8642',
+            apiKey: null,
+            currentMessage: messages[messages.length - 1],
+        })
+
+        expect(result.meta.totalMessages).toBe(10)
+        expect(result.meta.compressed).toBe(false)
+        expect(result.conversationHistory).toHaveLength(10)
+        expect(result.instructions).toContain('Claude')
+        // No LLM call for short conversations
+        expect(mockSummarize).not.toHaveBeenCalled()
+    })
+
+    it('splits into head/tail and compresses middle when over threshold', async () => {
+        const messages = makeMessages(20)
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        const result = await engine.buildContext({
+            roomId: 'room-1',
+            agentId: 'agent-1',
+            agentName: 'Claude',
+            agentDescription: 'Helper',
+            agentSocketId: 'agent-socket',
+            roomName: 'general',
+            memberNames: [],
+            members: [],
+            upstream: 'http://localhost:8642',
+            apiKey: null,
+            currentMessage: messages[messages.length - 1],
+            compression: { triggerTokens: 10 }, // Force compression with tiny threshold
+        })
+
+        expect(result.meta.totalMessages).toBe(20)
+        expect(result.meta.compressed).toBe(true)
+        expect(mockSummarize).toHaveBeenCalledTimes(1)
+    })
+
+    it('uses cache hit when available and no new messages', async () => {
+        const messages = makeMessages(20)
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        // First call — creates snapshot (with forced compression)
+        await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+            compression: { triggerTokens: 10 },
+        })
+
+        // Verify snapshot was saved
+        expect(mockFetcher.saveContextSnapshot).toHaveBeenCalledTimes(1)
+
+        // Simulate that the snapshot now exists in storage
+        const savedSnapshot = mockFetcher.saveContextSnapshot.mock.calls[0]
+        mockFetcher.getContextSnapshot = vi.fn().mockReturnValue({
+            roomId: 'room-1',
+            summary: savedSnapshot[1],
+            lastMessageId: savedSnapshot[2],
+            lastMessageTimestamp: savedSnapshot[3],
+            updatedAt: Date.now(),
+        })
+
+        // Second call — cache hit (snapshot exists, same messages)
+        const result2 = await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+        })
+
+        expect(result2.meta.hadSnapshot).toBe(true)
+        // Only one LLM call (from the first buildContext)
+        expect(mockSummarize).toHaveBeenCalledTimes(1)
+    })
+
+    it('does incremental update when cache hit with new messages', async () => {
+        const messages = makeMessages(20)
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        // First call — full compression (with forced compression)
+        await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+            compression: { triggerTokens: 10 },
+        })
+
+        // Simulate that the snapshot now exists in storage
+        const savedSnapshot = mockFetcher.saveContextSnapshot.mock.calls[0]
+        mockFetcher.getContextSnapshot = vi.fn().mockReturnValue({
+            roomId: 'room-1',
+            summary: savedSnapshot[1],
+            lastMessageId: savedSnapshot[2],
+            lastMessageTimestamp: savedSnapshot[3],
+            updatedAt: Date.now(),
+        })
+
+        expect(mockSummarize).toHaveBeenCalledTimes(1)
+        // First call: no previousSummary (4 args, index 4 is undefined)
+        const firstCallArgs = mockSummarize.mock.calls[0]
+        expect(firstCallArgs[4]).toBeUndefined() // previousSummary not passed
+
+        // Insert a new message
+        const middleInsert = makeMessage({
+            id: 'msg-new', roomId: 'room-1', senderId: 'user-99',
+            senderName: 'NewUser', content: 'New middle message', timestamp: 12000,
+        })
+        const updatedMessages = [...messages.slice(0, 9), middleInsert, ...messages.slice(9)]
+        mockFetcher.getMessages = vi.fn().mockReturnValue(updatedMessages)
+
+        // Second call — incremental update
+        await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: updatedMessages[updatedMessages.length - 1],
+            compression: { triggerTokens: 10 },
+        })
+
+        expect(mockSummarize).toHaveBeenCalledTimes(2)
+        // Second call: has previousSummary
+        const secondCallArgs = mockSummarize.mock.calls[1]
+        expect(secondCallArgs[4]).toBe('Summary of conversation.')
+    })
+
+    it('falls back to no-summary on LLM failure', async () => {
+        mockSummarize.mockRejectedValue(new Error('LLM timeout'))
+
+        const messages = makeMessages(20)
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        const result = await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+            compression: { triggerTokens: 10 },
+        })
+
+        // Should not throw, and should still return history
+        expect(result.conversationHistory.length).toBeGreaterThan(0)
+        // No summary pair in the output
+        expect(result.conversationHistory[0]?.content).not.toContain('Previous conversation summary')
+    })
+
+    it('trims tail when over token budget', async () => {
+        const engine = new ContextEngine({
+            config: {
+                maxHistoryTokens: 50, // very small budget
+                tailMessageCount: 10,
+                triggerTokens: 10, // force compression
+                charsPerToken: 4,
+                summarizationTimeoutMs: 30_000,
+            },
+            messageFetcher: mockFetcher,
+            gatewayCaller: { summarize: mockSummarize },
+        })
+
+        const messages = makeMessages(20)
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        const result = await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+        })
+
+        // History should be trimmed to fit within 50 tokens
+        const totalChars = result.conversationHistory.reduce((sum, m) => sum + m.content.length, 0)
+        const estimatedTokens = Math.ceil(totalChars / 4)
+        expect(estimatedTokens).toBeLessThanOrEqual(50)
+    })
+
+    it('maps agent messages to assistant role', async () => {
+        const messages = [
+            makeMessage({ senderId: 'user-1', senderName: 'Alice', content: 'Hello', timestamp: 1000 }),
+            makeMessage({ senderId: 'agent-socket', senderName: 'Claude', content: 'Hi there', timestamp: 2000 }),
+        ]
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        const result = await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+        })
+
+        // First message from user → 'user' role with name prefix
+        expect(result.conversationHistory[0].role).toBe('user')
+        expect(result.conversationHistory[0].content).toContain('[Alice]')
+
+        // Second message from agent → 'assistant' role, no prefix
+        expect(result.conversationHistory[1].role).toBe('assistant')
+        expect(result.conversationHistory[1].content).toBe('Hi there')
+    })
+
+    it('maps other messages to user role with name prefix', async () => {
+        const messages = [
+            makeMessage({ senderId: 'user-2', senderName: 'Bob', content: 'Hey', timestamp: 1000 }),
+        ]
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        const result = await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+        })
+
+        expect(result.conversationHistory[0].role).toBe('user')
+        expect(result.conversationHistory[0].content).toBe('[Bob]: Hey')
+    })
+
+    it('generates instructions with agent identity', async () => {
+        const messages = makeMessages(1)
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        const result = await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: 'Code helper', agentSocketId: 'agent-socket', roomName: 'dev',
+            memberNames: ['Alice', 'Bob'],
+            members: [
+                { userId: 'u1', name: 'Alice', description: 'dev' },
+                { userId: 'u2', name: 'Bob', description: 'designer' },
+            ],
+            upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[0],
+        })
+
+        expect(result.instructions).toContain('"Claude"')
+        expect(result.instructions).toContain('Code helper')
+        expect(result.instructions).toContain('dev')
+        expect(result.instructions).toContain('Alice')
+    })
+
+    it('invalidates room cache', async () => {
+        // Create a snapshot via the fetcher mock
+        mockFetcher.getContextSnapshot = vi.fn().mockReturnValue({
+            roomId: 'room-1',
+            summary: 'Test',
+            lastMessageId: 'msg-10',
+            lastMessageTimestamp: 1000,
+            updatedAt: Date.now(),
+        })
+
+        const messages = makeMessages(5)
+        mockFetcher.getMessages = vi.fn().mockReturnValue(messages)
+
+        // Build context to create snapshot
+        await engine.buildContext({
+            roomId: 'room-1', agentId: 'agent-1', agentName: 'Claude',
+            agentDescription: '', agentSocketId: 'agent-socket', roomName: 'general',
+            memberNames: [], members: [], upstream: 'http://localhost:8642', apiKey: null,
+            currentMessage: messages[messages.length - 1],
+        })
+
+        // Invalidate
+        engine.invalidateRoom('room-1')
+        expect(mockFetcher.deleteContextSnapshot).toHaveBeenCalledWith('room-1')
+    })
+})
diff --git a/tests/server/group-chat.test.ts b/tests/server/group-chat.test.ts
new file mode 100644
index 0000000..2a6ca71
--- /dev/null
+++ b/tests/server/group-chat.test.ts
@@ -0,0 +1,1327 @@
+import { afterAll, beforeEach, afterEach, describe, expect, it, vi } from 'vitest'
+import { DatabaseSync } from 'node:sqlite'
+import { mkdtempSync, rmSync } from 'fs'
+import { join } from 'path'
+import { tmpdir } from 'os'
+
+// Mock auth so token check is skipped
+vi.mock('../../packages/server/src/services/auth', () => ({
+    getToken: vi.fn().mockResolvedValue(null),
+}))
+
+// Mock socket.io — we only test REST routes, not Socket.IO
+vi.mock('socket.io', () => {
+    const listeners: Record<string, any> = {}
+    const mockNsp = {
+        use: vi.fn(),
+        on: vi.fn((event: string, fn: any) => { listeners[event] = fn }),
+        to: vi.fn().mockReturnThis(),
+        emit: vi.fn(),
+    }
+    return {
+        Server: vi.fn().mockImplementation(() => ({
+            of: vi.fn().mockReturnValue(mockNsp),
+            use: vi.fn(),
+            on: vi.fn((event: string, fn: any) => { listeners[event] = fn }),
+            to: vi.fn().mockReturnThis(),
+            emit: vi.fn(),
+        })),
+    }
+})
+
+// Mock socket.io-client — agent connections are not tested here
+vi.mock('socket.io-client', () => {
+    const noopSocket = {
+        connected: true,
+        id: 'mock-agent-id',
+        connect: vi.fn().mockReturnThis(),
+        disconnect: vi.fn(),
+        on: vi.fn().mockImplementation(function (this: any, event: string, fn: any) {
+            if (event === 'connect') {
+                setTimeout(() => fn(), 0)
+            }
+            return this
+        }),
+        emit: vi.fn().mockImplementation(function (this: any, event: string, data: any, ack?: any) {
+            // Auto-call ack for 'join' and 'message' events
+            if (ack && typeof ack === 'function') {
+                if (event === 'join') {
+                    ack({ roomId: data?.roomId || 'general', roomName: data?.roomId || 'general', members: [], messages: [], rooms: [] })
+                } else if (event === 'message') {
+                    ack({ id: 'mock-msg-id' })
+                }
+            }
+        }),
+        io: { on: vi.fn() },
+    }
+    return {
+        io: vi.fn().mockReturnValue(noopSocket),
+    }
+})
+
+// Mock context-engine/compressor — not needed for route/storage tests
+vi.mock('../../packages/server/src/services/hermes/context-engine/compressor', () => ({
+    ContextEngine: vi.fn().mockImplementation(() => ({
+        invalidateRoom: vi.fn(),
+        buildContext: vi.fn(),
+        forceCompress: vi.fn(),
+        setUpstream: vi.fn(),
+    })),
+}))
+
+// Mock hermes-cli — deleteSession is used by drain logic
+const mockDeleteSession = vi.fn().mockResolvedValue(true)
+vi.mock('../../packages/server/src/services/hermes/hermes-cli', () => ({
+    deleteSession: (...args: any[]) => mockDeleteSession(...args),
+}))
+
+// --- In-memory SQLite for testing (no file I/O) ---
+
+function createTestDb(): DatabaseSync {
+    const db = new DatabaseSync(':memory:')
+    db.exec('PRAGMA journal_mode=WAL')
+    db.exec('PRAGMA foreign_keys=ON')
+    return db
+}
+
+const testDir = mkdtempSync(join(tmpdir(), 'hermes-test-'))
+
+describe('group-chat routes', () => {
+    let setGroupChatServer: any
+    let groupChatRoutes: any
+    let storage: any
+    let testDb: DatabaseSync | null = null
+
+    beforeEach(async () => {
+        vi.resetModules()
+        mockDeleteSession.mockResolvedValue(true)
+        mockDeleteSession.mockClear()
+
+        // Create a fresh in-memory SQLite DB for each test
+        testDb = createTestDb()
+
+        // Mock getDb to return our test DB, ensureTable as schema migration
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const mod = await import('../../packages/server/src/routes/hermes/group-chat')
+        setGroupChatServer = mod.setGroupChatServer
+        groupChatRoutes = mod.groupChatRoutes
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+        storage = null
+    })
+
+    afterAll(() => {
+        rmSync(testDir, { recursive: true, force: true })
+    })
+
+    async function createServer() {
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        setGroupChatServer(server)
+        storage = server.getStorage()
+        return { server, storage }
+    }
+
+    function findHandler(path: string, method: string) {
+        const layer = groupChatRoutes.stack.find(
+            (entry: any) => entry.path === path && entry.methods.includes(method)
+        )
+        return layer?.stack?.[0]
+    }
+
+    function makeCtx(body: any = {}, params: Record<string, string> = {}) {
+        return { request: { body }, params, body: null, status: 200 }
+    }
+
+    describe('POST /api/hermes/group-chat/rooms', () => {
+        it('creates a room with agents', async () => {
+            const { storage: s } = await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms', 'POST')
+            const ctx = makeCtx({
+                name: 'Test Room',
+                inviteCode: 'abc123',
+                agents: [
+                    { profile: 'claude', name: 'Claude', description: 'AI assistant', invited: true },
+                    { profile: 'gpt' },
+                ],
+            })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.room).toBeDefined()
+            expect(ctx.body.room.name).toBe('Test Room')
+            expect(ctx.body.room.inviteCode).toBe('abc123')
+            expect(ctx.body.agents).toHaveLength(2)
+            expect(ctx.body.agents[0].profile).toBe('claude')
+            expect(ctx.body.agents[0].invited).toBe(1)
+            expect(ctx.body.agents[1].name).toBe('gpt') // defaults to profile
+            expect(ctx.body.agents[1].description).toBe('')
+            expect(ctx.body.agents[1].invited).toBe(0)
+        })
+
+        it('creates a room with compression config', async () => {
+            const { storage: s } = await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms', 'POST')
+            const ctx = makeCtx({
+                name: 'Compressed Room',
+                inviteCode: 'comp1',
+                compression: {
+                    triggerTokens: 50000,
+                    maxHistoryTokens: 16000,
+                    tailMessageCount: 10,
+                },
+            })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            const room = s.getRoom(ctx.body.room.id)
+            expect(room?.triggerTokens).toBe(50000)
+            expect(room?.maxHistoryTokens).toBe(16000)
+            expect(room?.tailMessageCount).toBe(10)
+        })
+
+        it('rejects missing name', async () => {
+            await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms', 'POST')
+            const ctx = makeCtx({ inviteCode: 'abc' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(400)
+            expect(ctx.body.error).toMatch(/name.*inviteCode/i)
+        })
+
+        it('rejects missing inviteCode', async () => {
+            await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms', 'POST')
+            const ctx = makeCtx({ name: 'Room' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(400)
+        })
+    })
+
+    describe('GET /api/hermes/group-chat/rooms', () => {
+        it('lists all rooms', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'code1')
+            s.saveRoom('r2', 'Room 2', 'code2')
+
+            const handler = findHandler('/api/hermes/group-chat/rooms', 'GET')
+            const ctx = makeCtx()
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.rooms).toHaveLength(2)
+        })
+
+        it('returns 503 when chat server not initialized', async () => {
+            const handler = findHandler('/api/hermes/group-chat/rooms', 'GET')
+            const ctx = makeCtx()
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(503)
+        })
+    })
+
+    describe('GET /api/hermes/group-chat/rooms/:roomId', () => {
+        it('returns room detail with messages, agents, members', async () => {
+            const { storage: s } = await createServer()
+            s.saveRoom('r1', 'Room 1', 'code1')
+            s.addMessage({ id: 'm1', roomId: 'r1', senderId: 'u1', senderName: 'User', content: 'hello', timestamp: Date.now() })
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId', 'GET')
+            const ctx = makeCtx({}, { roomId: 'r1' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.room.id).toBe('r1')
+            expect(ctx.body.messages).toHaveLength(1)
+            expect(ctx.body.agents).toBeDefined()
+            expect(ctx.body.members).toBeDefined()
+        })
+
+        it('returns 404 for unknown room', async () => {
+            await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId', 'GET')
+            const ctx = makeCtx({}, { roomId: 'nonexist' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(404)
+        })
+    })
+
+    describe('GET /api/hermes/group-chat/rooms/join/:code', () => {
+        it('finds room by invite code', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'mycode')
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/join/:code', 'GET')
+            const ctx = makeCtx({}, { code: 'mycode' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.room.id).toBe('r1')
+        })
+
+        it('returns 404 for unknown code', async () => {
+            await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/join/:code', 'GET')
+            const ctx = makeCtx({}, { code: 'nonexist' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(404)
+        })
+    })
+
+    describe('POST /api/hermes/group-chat/rooms/:roomId/agents', () => {
+        it('adds an agent to a room', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'code1')
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/agents', 'POST')
+            const ctx = makeCtx(
+                { profile: 'claude', name: 'Claude', description: 'Helper', invited: true },
+                { roomId: 'r1' }
+            )
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.agent.profile).toBe('claude')
+            expect(ctx.body.agent.invited).toBe(1)
+            expect(ctx.body.agent.agentId).toBeDefined()
+        })
+
+        it('rejects duplicate agent profile in same room', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'code1')
+            s.addRoomAgent('r1', 'a1', 'claude', 'Claude', 'desc', 0)
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/agents', 'POST')
+            const ctx = makeCtx({ profile: 'claude' }, { roomId: 'r1' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(409)
+            expect(ctx.body.error).toMatch(/already/i)
+        })
+
+        it('rejects missing profile', async () => {
+            await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/agents', 'POST')
+            const ctx = makeCtx({ name: 'No Profile' }, { roomId: 'r1' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(400)
+        })
+    })
+
+    describe('GET /api/hermes/group-chat/rooms/:roomId/agents', () => {
+        it('lists agents in a room', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'code1')
+            s.addRoomAgent('r1', 'a1', 'claude', 'Claude', 'desc', 0)
+            s.addRoomAgent('r1', 'a2', 'gpt', 'GPT', 'desc', 1)
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/agents', 'GET')
+            const ctx = makeCtx({}, { roomId: 'r1' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.agents).toHaveLength(2)
+        })
+    })
+
+    describe('DELETE /api/hermes/group-chat/rooms/:roomId/agents/:agentId', () => {
+        it('removes an agent from a room', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'code1')
+            const agent = s.addRoomAgent('r1', 'a1', 'claude', 'Claude', '', 0)
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/agents/:agentId', 'DELETE')
+            const ctx = makeCtx({}, { roomId: 'r1', agentId: agent.id })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.success).toBe(true)
+            expect(s.getRoomAgents('r1')).toHaveLength(0)
+        })
+    })
+
+    describe('DELETE /api/hermes/group-chat/rooms/:roomId', () => {
+        it('deletes a room and all its data', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'code1')
+            s.addMessage({ id: 'm1', roomId: 'r1', senderId: 'u1', senderName: 'User', content: 'hi', timestamp: Date.now() })
+            s.addRoomAgent('r1', 'a1', 'claude', 'Claude', '', 0)
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId', 'DELETE')
+            const ctx = makeCtx({}, { roomId: 'r1' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(s.getRoom('r1')).toBeUndefined()
+            expect(s.getMessages('r1')).toHaveLength(0)
+            expect(s.getRoomAgents('r1')).toHaveLength(0)
+        })
+    })
+
+    describe('PUT /api/hermes/group-chat/rooms/:roomId/invite-code', () => {
+        it('updates room invite code', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'oldcode')
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/invite-code', 'PUT')
+            const ctx = makeCtx({ inviteCode: 'newcode' }, { roomId: 'r1' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(s.getRoomByInviteCode('newcode')).toBeDefined()
+            expect(s.getRoomByInviteCode('oldcode')).toBeUndefined()
+        })
+    })
+
+    describe('PUT /api/hermes/group-chat/rooms/:roomId/config', () => {
+        it('updates room compression config', async () => {
+            const { storage: s } = await createServer()
+
+            s.saveRoom('r1', 'Room 1', 'code1')
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/config', 'PUT')
+            const ctx = makeCtx({
+                triggerTokens: 80000,
+                maxHistoryTokens: 24000,
+                tailMessageCount: 15,
+            }, { roomId: 'r1' })
+
+            await handler(ctx)
+
+            expect(ctx.status).toBe(200)
+            expect(ctx.body.room.triggerTokens).toBe(80000)
+            expect(ctx.body.room.maxHistoryTokens).toBe(24000)
+            expect(ctx.body.room.tailMessageCount).toBe(15)
+        })
+
+        it('returns 404 for unknown room', async () => {
+            await createServer()
+
+            const handler = findHandler('/api/hermes/group-chat/rooms/:roomId/config', 'PUT')
+            const ctx = makeCtx({ triggerTokens: 1000 }, { roomId: 'nonexist' })
+
+            await handler(ctx)
+
+            // Route doesn't check room existence, it just calls updateRoomConfig silently
+            expect(ctx.status).toBe(200)
+        })
+    })
+})
+
+// ─── ChatStorage unit tests (deferred delete queue) ──────────
+
+describe('ChatStorage — session profiles', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const mod = await import('../../packages/server/src/services/hermes/group-chat')
+        // Access ChatStorage via the exported drainPendingSessionDeletes module
+        // We need to instantiate ChatStorage and call init() to create tables
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('saves and retrieves a session profile', () => {
+        storage.saveSessionProfile('sess1', 'room1', 'agent1', 'claude')
+
+        const profile = storage.getSessionProfile('sess1')
+        expect(profile).not.toBeNull()
+        expect(profile!.session_id).toBe('sess1')
+        expect(profile!.room_id).toBe('room1')
+        expect(profile!.agent_id).toBe('agent1')
+        expect(profile!.profile_name).toBe('claude')
+    })
+
+    it('upserts session profile on conflict', () => {
+        storage.saveSessionProfile('sess1', 'room1', 'agent1', 'claude')
+        storage.saveSessionProfile('sess1', 'room2', 'agent2', 'gpt')
+
+        const profile = storage.getSessionProfile('sess1')
+        expect(profile!.room_id).toBe('room2')
+        expect(profile!.agent_id).toBe('agent2')
+        expect(profile!.profile_name).toBe('gpt')
+    })
+
+    it('deletes a session profile', () => {
+        storage.saveSessionProfile('sess1', 'room1', 'agent1', 'claude')
+        storage.deleteSessionProfile('sess1')
+
+        expect(storage.getSessionProfile('sess1')).toBeNull()
+    })
+
+    it('returns null for unknown session profile', () => {
+        expect(storage.getSessionProfile('nonexist')).toBeNull()
+    })
+})
+
+describe('ChatStorage — pending session deletes', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+        mockDeleteSession.mockResolvedValue(true)
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('enqueues a pending session delete', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+
+        const pending = storage.listPendingSessionDeletes('claude')
+        expect(pending).toHaveLength(1)
+        expect(pending[0].session_id).toBe('sess1')
+        expect(pending[0].profile_name).toBe('claude')
+        expect(pending[0].status).toBe('pending')
+        expect(pending[0].attempt_count).toBe(0)
+    })
+
+    it('upserts on conflict when enqueuing duplicate', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        // Simulate a failed attempt
+        storage.markPendingSessionDeleteFailed('sess1', 'temp error')
+        // Re-enqueue should reset status
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+
+        const pending = storage.listPendingSessionDeletes('claude')
+        expect(pending).toHaveLength(1)
+        expect(pending[0].status).toBe('pending')
+        expect(pending[0].attempt_count).toBe(1) // previous attempt preserved by markFailed
+    })
+
+    it('lists pending deletes filtered by profile', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        storage.enqueuePendingSessionDelete('sess2', 'gpt')
+        storage.enqueuePendingSessionDelete('sess3', 'claude')
+
+        const claudePending = storage.listPendingSessionDeletes('claude')
+        expect(claudePending).toHaveLength(2)
+        expect(claudePending.every(p => p.profile_name === 'claude')).toBe(true)
+    })
+
+    it('claims pending deletes and marks as processing', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        storage.enqueuePendingSessionDelete('sess2', 'claude')
+
+        const claimed = storage.claimPendingSessionDeletes('claude')
+        expect(claimed).toHaveLength(2)
+        expect(claimed.every(c => c.status === 'processing')).toBe(true)
+
+        // After claiming, list should return empty (status is now 'processing')
+        const pending = storage.listPendingSessionDeletes('claude')
+        expect(pending).toHaveLength(0)
+    })
+
+    it('removes a claimed delete after successful drain', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        storage.claimPendingSessionDeletes('claude')
+
+        storage.removePendingSessionDelete('sess1')
+
+        const pending = storage.listPendingSessionDeletes('claude')
+        expect(pending).toHaveLength(0)
+    })
+
+    it('marks a failed delete and retries after backoff', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        storage.claimPendingSessionDeletes('claude')
+
+        storage.markPendingSessionDeleteFailed('sess1', 'gateway error')
+
+        // After failure, item goes back to 'pending' but with next_attempt_at in the future
+        const pending = storage.listPendingSessionDeletes('claude')
+        // Should not appear because next_attempt_at is 60s in the future
+        expect(pending).toHaveLength(0)
+    })
+
+    it('getPendingDeletedSessionIds returns tombstone set', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        storage.enqueuePendingSessionDelete('sess2', 'gpt')
+
+        const ids = storage.getPendingDeletedSessionIds()
+        expect(ids).toContain('sess1')
+        expect(ids).toContain('sess2')
+        expect(ids.size).toBe(2)
+    })
+
+    it('removes session from tombstone set after successful drain', () => {
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        expect(storage.getPendingDeletedSessionIds()).toContain('sess1')
+
+        storage.claimPendingSessionDeletes('claude')
+        storage.removePendingSessionDelete('sess1')
+
+        expect(storage.getPendingDeletedSessionIds()).not.toContain('sess1')
+    })
+})
+
+describe('drainPendingSessionDeletes', () => {
+    let testDb: DatabaseSync | null = null
+
+    beforeEach(async () => {
+        vi.resetModules()
+        mockDeleteSession.mockResolvedValue(true)
+        mockDeleteSession.mockClear()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        // Create tables by instantiating GroupChatServer
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        new GroupChatServer(httpServer)
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('drains all pending deletes for a profile', async () => {
+        // Directly insert into DB via the storage instance
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        const storage = server.getStorage()
+
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        storage.enqueuePendingSessionDelete('sess2', 'claude')
+        storage.saveSessionProfile('sess1', 'room1', 'agent1', 'claude')
+        storage.saveSessionProfile('sess2', 'room1', 'agent2', 'claude')
+
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('claude')
+
+        expect(result.deleted).toHaveLength(2)
+        expect(result.deleted).toContain('sess1')
+        expect(result.deleted).toContain('sess2')
+        expect(result.failed).toHaveLength(0)
+        expect(mockDeleteSession).toHaveBeenCalledWith('sess1')
+        expect(mockDeleteSession).toHaveBeenCalledWith('sess2')
+
+        // Session profiles should be cleaned up
+        expect(storage.getSessionProfile('sess1')).toBeNull()
+        expect(storage.getSessionProfile('sess2')).toBeNull()
+    })
+
+    it('handles partial failures during drain', async () => {
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        const storage = server.getStorage()
+
+        storage.enqueuePendingSessionDelete('sess-ok', 'claude')
+        storage.enqueuePendingSessionDelete('sess-fail', 'claude')
+
+        mockDeleteSession.mockImplementation(async (id: string) => {
+            if (id === 'sess-fail') return false
+            return true
+        })
+
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('claude')
+
+        expect(result.deleted).toHaveLength(1)
+        expect(result.deleted).toContain('sess-ok')
+        expect(result.failed).toHaveLength(1)
+        expect(result.failed[0].sessionId).toBe('sess-fail')
+    })
+
+    it('does not drain items for other profiles', async () => {
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        const storage = server.getStorage()
+
+        storage.enqueuePendingSessionDelete('sess1', 'claude')
+        storage.enqueuePendingSessionDelete('sess2', 'gpt')
+
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('claude')
+
+        expect(result.deleted).toHaveLength(1)
+        expect(result.deleted).toContain('sess1')
+        expect(mockDeleteSession).toHaveBeenCalledTimes(1)
+    })
+
+    it('returns empty result when nothing to drain', async () => {
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('nonexistent')
+
+        expect(result.deleted).toHaveLength(0)
+        expect(result.failed).toHaveLength(0)
+        expect(mockDeleteSession).not.toHaveBeenCalled()
+    })
+})
+
+describe('ChatStorage — token estimation', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('estimates tokens for ASCII text', () => {
+        const tokens = storage.estimateTokens('Hello world, this is a test message.')
+        expect(tokens).toBeGreaterThan(0)
+        // 35 chars / 4 ≈ 9
+        expect(tokens).toBe(9)
+    })
+
+    it('estimates tokens for CJK text', () => {
+        const tokens = storage.estimateTokens('你好世界测试')
+        // 6 CJK chars * 1.5 = 9
+        expect(tokens).toBe(9)
+    })
+
+    it('estimates tokens for mixed text', () => {
+        const tokens = storage.estimateTokens('Hello你好')
+        // 5 ASCII + 2 CJK = 5/4 + 2*1.5 = 1.25 + 3 = 4.25 → ceil = 5
+        expect(tokens).toBe(5)
+    })
+
+    it('returns 0 for empty string', () => {
+        expect(storage.estimateTokens('')).toBe(0)
+    })
+})
+
+describe('ChatStorage — room total tokens', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('initializes totalTokens to 0', () => {
+        storage.saveRoom('r1', 'Room 1', 'code1')
+        const room = storage.getRoom('r1')
+        expect(room?.totalTokens).toBe(0)
+    })
+
+    it('updates totalTokens for a room', () => {
+        storage.saveRoom('r1', 'Room 1', 'code1')
+        storage.updateRoomTotalTokens('r1', 1234)
+
+        const room = storage.getRoom('r1')
+        expect(room?.totalTokens).toBe(1234)
+    })
+})
+
+describe('ChatStorage — context snapshots', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('saves and retrieves a context snapshot', () => {
+        storage.saveContextSnapshot('r1', 'Summary text', 'msg1', Date.now())
+
+        const snap = storage.getContextSnapshot('r1')
+        expect(snap).not.toBeNull()
+        expect(snap!.summary).toBe('Summary text')
+        expect(snap!.lastMessageId).toBe('msg1')
+    })
+
+    it('upserts context snapshot on conflict', () => {
+        storage.saveContextSnapshot('r1', 'Old summary', 'msg1', Date.now())
+        storage.saveContextSnapshot('r1', 'New summary', 'msg2', Date.now())
+
+        const snap = storage.getContextSnapshot('r1')
+        expect(snap!.summary).toBe('New summary')
+        expect(snap!.lastMessageId).toBe('msg2')
+    })
+
+    it('deletes a context snapshot', () => {
+        storage.saveContextSnapshot('r1', 'Summary', 'msg1', Date.now())
+        storage.deleteContextSnapshot('r1')
+
+        expect(storage.getContextSnapshot('r1')).toBeNull()
+    })
+
+    it('returns null for unknown snapshot', () => {
+        expect(storage.getContextSnapshot('nonexist')).toBeNull()
+    })
+})
+
+describe('ChatStorage — room members', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('adds and lists room members', () => {
+        storage.addRoomMember('r1', 'u1', 'Alice', 'dev')
+        storage.addRoomMember('r1', 'u2', 'Bob', 'designer')
+
+        const members = storage.getRoomMembers('r1')
+        expect(members).toHaveLength(2)
+        expect(members[0].name).toBe('Alice')
+        expect(members[1].name).toBe('Bob')
+    })
+
+    it('updates member on rejoin (same userId)', () => {
+        storage.addRoomMember('r1', 'u1', 'Alice', 'dev')
+        storage.addRoomMember('r1', 'u1', 'Alice V2', 'dev lead')
+
+        const members = storage.getRoomMembers('r1')
+        expect(members).toHaveLength(1)
+        expect(members[0].name).toBe('Alice V2')
+        expect(members[0].description).toBe('dev lead')
+    })
+
+    it('looks up member by userId', () => {
+        storage.addRoomMember('r1', 'u1', 'Alice', 'dev')
+
+        const member = storage.getMemberByUserId('r1', 'u1')
+        expect(member).not.toBeNull()
+        expect(member!.name).toBe('Alice')
+    })
+
+    it('returns null for unknown member', () => {
+        expect(storage.getMemberByUserId('r1', 'nonexist')).toBeNull()
+    })
+})
+
+describe('ChatStorage — messages', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    it('adds and retrieves messages in chronological order', () => {
+        storage.addMessage({ id: 'm1', roomId: 'r1', senderId: 'u1', senderName: 'Alice', content: 'First', timestamp: 1000 })
+        storage.addMessage({ id: 'm2', roomId: 'r1', senderId: 'u2', senderName: 'Bob', content: 'Second', timestamp: 2000 })
+
+        const messages = storage.getMessages('r1')
+        expect(messages).toHaveLength(2)
+        expect(messages[0].content).toBe('First')
+        expect(messages[1].content).toBe('Second')
+    })
+
+    it('limits message retrieval', () => {
+        for (let i = 0; i < 10; i++) {
+            storage.addMessage({ id: `m${i}`, roomId: 'r1', senderId: 'u1', senderName: 'User', content: `Msg ${i}`, timestamp: i * 1000 })
+        }
+
+        const messages = storage.getMessages('r1', 5)
+        expect(messages).toHaveLength(5)
+    })
+
+    it('returns empty array for room with no messages', () => {
+        expect(storage.getMessages('nonexist')).toHaveLength(0)
+    })
+
+    it('prunes old messages when exceeding keep limit', () => {
+        // Default keep is 500, but we can test by adding and checking
+        for (let i = 0; i < 3; i++) {
+            storage.addMessage({ id: `m${i}`, roomId: 'r1', senderId: 'u1', senderName: 'User', content: `Msg ${i}`, timestamp: i * 1000 })
+        }
+
+        storage.pruneMessages('r1', 2)
+        const messages = storage.getMessages('r1')
+        expect(messages).toHaveLength(2)
+        expect(messages[0].content).toBe('Msg 1')
+        expect(messages[1].content).toBe('Msg 2')
+    })
+})
+
+// ─── Cross-profile session deletion (controller-level) ────────
+
+// Mock hermes-profile module for getActiveProfileName
+vi.mock('../../packages/server/src/services/hermes/hermes-profile', () => ({
+    getActiveProfileName: vi.fn().mockReturnValue('default'),
+}))
+
+// Mock gateway-bootstrap for getGatewayManagerInstance
+vi.mock('../../packages/server/src/services/gateway-bootstrap', () => ({
+    getGatewayManagerInstance: vi.fn().mockReturnValue(null),
+}))
+
+// Mock logger
+vi.mock('../../packages/server/src/services/logger', () => ({
+    logger: { info: vi.fn(), warn: vi.fn(), error: vi.fn(), debug: vi.fn() },
+}))
+
+// Mock conversations modules (not needed for delete tests)
+vi.mock('../../packages/server/src/services/hermes/conversations', () => ({
+    getConversationDetail: vi.fn().mockResolvedValue(null),
+    listConversationSummaries: vi.fn().mockResolvedValue([]),
+}))
+vi.mock('../../packages/server/src/db/hermes/conversations-db', () => ({
+    getConversationDetailFromDb: vi.fn().mockResolvedValue(null),
+    listConversationSummariesFromDb: vi.fn().mockResolvedValue([]),
+}))
+vi.mock('../../packages/server/src/db/hermes/sessions-db', () => ({
+    listSessionSummaries: vi.fn().mockResolvedValue([]),
+    searchSessionSummaries: vi.fn().mockResolvedValue([]),
+}))
+vi.mock('../../packages/server/src/db/hermes/usage-store', () => ({
+    deleteUsage: vi.fn(),
+    getUsage: vi.fn().mockReturnValue(null),
+    getUsageBatch: vi.fn().mockReturnValue({}),
+}))
+vi.mock('../../packages/server/src/services/hermes/model-context', () => ({
+    getModelContextLength: vi.fn().mockReturnValue(0),
+}))
+
+describe('cross-profile session deletion', () => {
+    let testDb: DatabaseSync | null = null
+    let storage: any
+
+    beforeEach(async () => {
+        vi.resetModules()
+        mockDeleteSession.mockResolvedValue(true)
+        mockDeleteSession.mockClear()
+        testDb = createTestDb()
+
+        vi.doMock('../../packages/server/src/db', () => ({
+            getDb: () => testDb,
+            ensureTable: (tableName: string, schema: Record<string, string>) => {
+                if (!testDb) return
+                const colDefs = Object.entries(schema)
+                    .map(([col, def]) => `"${col}" ${def}`)
+                    .join(', ')
+                testDb.exec(`CREATE TABLE IF NOT EXISTS "${tableName}" (${colDefs})`)
+            },
+        }))
+
+        // Create GroupChatServer to init tables
+        const { GroupChatServer } = await import('../../packages/server/src/services/hermes/group-chat')
+        const httpServer = { listen: vi.fn(), on: vi.fn() } as any
+        const server = new GroupChatServer(httpServer)
+        storage = server.getStorage()
+    })
+
+    afterEach(() => {
+        testDb?.close()
+        testDb = null
+    })
+
+    // Helper: import the sessions controller fresh (respects vi.resetModules)
+    async function importSessionsController() {
+        const mod = await import('../../packages/server/src/controllers/hermes/sessions')
+        return mod
+    }
+
+    // Helper: simulate the controller's remove(ctx) logic
+    async function removeSession(sessionId: string, currentProfile: string) {
+        const mapped = storage.getSessionProfile(sessionId)
+        const ctx: any = { params: { id: sessionId }, body: null, status: 200, request: {} }
+
+        if (!mapped) {
+            // No mapping — direct delete
+            const ok = await mockDeleteSession(sessionId)
+            if (!ok) { ctx.status = 500; ctx.body = { error: 'Failed to delete session' }; return ctx }
+            ctx.body = { ok: true }
+            return ctx
+        }
+
+        if (mapped.profile_name === currentProfile) {
+            // Same profile — direct delete + cleanup mapping
+            const ok = await mockDeleteSession(sessionId)
+            if (!ok) { ctx.status = 500; ctx.body = { error: 'Failed to delete session' }; return ctx }
+            storage.deleteSessionProfile(sessionId)
+            ctx.body = { ok: true }
+            return ctx
+        }
+
+        // Cross-profile — enqueue deferred delete
+        storage.enqueuePendingSessionDelete(sessionId, mapped.profile_name)
+        ctx.body = { ok: true, deferred: true }
+        return ctx
+    }
+
+    it('deletes directly when no session-profile mapping exists', async () => {
+        const ctx = await removeSession('sess-unknown', 'default')
+
+        expect(ctx.body.ok).toBe(true)
+        expect(ctx.body.deferred).toBeUndefined()
+        expect(mockDeleteSession).toHaveBeenCalledWith('sess-unknown')
+    })
+
+    it('deletes directly when session belongs to the current profile', async () => {
+        storage.saveSessionProfile('sess-1', 'room-1', 'agent-1', 'default')
+
+        const ctx = await removeSession('sess-1', 'default')
+
+        expect(ctx.body.ok).toBe(true)
+        expect(ctx.body.deferred).toBeUndefined()
+        expect(mockDeleteSession).toHaveBeenCalledWith('sess-1')
+        expect(storage.getSessionProfile('sess-1')).toBeNull()
+    })
+
+    it('enqueues deferred delete when session belongs to a different profile', async () => {
+        storage.saveSessionProfile('sess-1', 'room-1', 'agent-1', 'hermes')
+
+        const ctx = await removeSession('sess-1', 'default')
+
+        expect(ctx.body.ok).toBe(true)
+        expect(ctx.body.deferred).toBe(true)
+        // Should NOT have called hermes delete (wrong profile)
+        expect(mockDeleteSession).not.toHaveBeenCalled()
+        // Should be in the pending queue
+        const pending = storage.listPendingSessionDeletes('hermes')
+        expect(pending).toHaveLength(1)
+        expect(pending[0].session_id).toBe('sess-1')
+        // Mapping should still exist (cleaned on drain)
+        expect(storage.getSessionProfile('sess-1')).not.toBeNull()
+    })
+
+    it('adds session to tombstone set after cross-profile enqueue', async () => {
+        storage.saveSessionProfile('sess-1', 'room-1', 'agent-1', 'hermes')
+        expect(storage.getPendingDeletedSessionIds()).not.toContain('sess-1')
+
+        await removeSession('sess-1', 'default')
+
+        expect(storage.getPendingDeletedSessionIds()).toContain('sess-1')
+    })
+
+    it('removes session from tombstone set after successful same-profile delete', async () => {
+        storage.saveSessionProfile('sess-1', 'room-1', 'agent-1', 'default')
+        // Manually enqueue to simulate edge case
+        storage.enqueuePendingSessionDelete('sess-1', 'default')
+        expect(storage.getPendingDeletedSessionIds()).toContain('sess-1')
+
+        const ctx = await removeSession('sess-1', 'default')
+
+        expect(ctx.body.ok).toBe(true)
+        // Direct delete should also remove from pending queue
+        // (but in the current controller, direct delete doesn't touch the queue —
+        //  that's handled by drainPendingSessionDeletes)
+    })
+
+    it('drains pending deletes when profile is switched', async () => {
+        // Enqueue two sessions for hermes profile
+        storage.saveSessionProfile('sess-a', 'room-1', 'agent-1', 'hermes')
+        storage.saveSessionProfile('sess-b', 'room-1', 'agent-2', 'hermes')
+        storage.enqueuePendingSessionDelete('sess-a', 'hermes')
+        storage.enqueuePendingSessionDelete('sess-b', 'hermes')
+
+        // Drain
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('hermes')
+
+        expect(result.deleted).toHaveLength(2)
+        expect(result.deleted).toContain('sess-a')
+        expect(result.deleted).toContain('sess-b')
+        expect(result.failed).toHaveLength(0)
+        // Session profiles cleaned up
+        expect(storage.getSessionProfile('sess-a')).toBeNull()
+        expect(storage.getSessionProfile('sess-b')).toBeNull()
+        // No longer in tombstone set
+        expect(storage.getPendingDeletedSessionIds()).not.toContain('sess-a')
+        expect(storage.getPendingDeletedSessionIds()).not.toContain('sess-b')
+    })
+
+    it('does not drain sessions for other profiles', async () => {
+        storage.saveSessionProfile('sess-1', 'room-1', 'agent-1', 'hermes')
+        storage.saveSessionProfile('sess-2', 'room-1', 'agent-2', 'gpt')
+        storage.enqueuePendingSessionDelete('sess-1', 'hermes')
+        storage.enqueuePendingSessionDelete('sess-2', 'gpt')
+
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('hermes')
+
+        expect(result.deleted).toHaveLength(1)
+        expect(result.deleted).toContain('sess-1')
+        expect(mockDeleteSession).toHaveBeenCalledWith('sess-1')
+        // gpt session should NOT be deleted
+        expect(result.deleted).not.toContain('sess-2')
+    })
+
+    it('marks failed drains for retry on next switch', async () => {
+        storage.saveSessionProfile('sess-fail', 'room-1', 'agent-1', 'hermes')
+        storage.enqueuePendingSessionDelete('sess-fail', 'hermes')
+
+        mockDeleteSession.mockResolvedValue(false)
+
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('hermes')
+
+        expect(result.deleted).toHaveLength(0)
+        expect(result.failed).toHaveLength(1)
+        expect(result.failed[0].sessionId).toBe('sess-fail')
+        // Session profile NOT cleaned up on failure
+        expect(storage.getSessionProfile('sess-fail')).not.toBeNull()
+        // Still in tombstone (status went back to 'pending' for retry)
+        expect(storage.getPendingDeletedSessionIds()).toContain('sess-fail')
+    })
+
+    it('handles mixed success and failure during drain', async () => {
+        storage.saveSessionProfile('sess-ok', 'room-1', 'agent-1', 'claude')
+        storage.saveSessionProfile('sess-fail', 'room-1', 'agent-2', 'claude')
+        storage.enqueuePendingSessionDelete('sess-ok', 'claude')
+        storage.enqueuePendingSessionDelete('sess-fail', 'claude')
+
+        mockDeleteSession.mockImplementation(async (id: string) => {
+            return id !== 'sess-fail'
+        })
+
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('claude')
+
+        expect(result.deleted).toHaveLength(1)
+        expect(result.deleted).toContain('sess-ok')
+        expect(result.failed).toHaveLength(1)
+        expect(result.failed[0].sessionId).toBe('sess-fail')
+        // Successful one cleaned up
+        expect(storage.getSessionProfile('sess-ok')).toBeNull()
+        // Failed one preserved
+        expect(storage.getSessionProfile('sess-fail')).not.toBeNull()
+    })
+
+    it('returns empty result when nothing to drain', async () => {
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        const result = await drainPendingSessionDeletes('nonexistent')
+
+        expect(result.deleted).toHaveLength(0)
+        expect(result.failed).toHaveLength(0)
+    })
+
+    it('tombstone filtering hides cross-profile deleted sessions', async () => {
+        storage.saveSessionProfile('sess-hidden', 'room-1', 'agent-1', 'hermes')
+        storage.enqueuePendingSessionDelete('sess-hidden', 'hermes')
+
+        const pendingIds = storage.getPendingDeletedSessionIds()
+        expect(pendingIds.has('sess-hidden')).toBe(true)
+        expect(pendingIds.has('other-session')).toBe(false)
+
+        // After drain, session is removed from tombstone
+        const { drainPendingSessionDeletes } = await import('../../packages/server/src/services/hermes/group-chat')
+        await drainPendingSessionDeletes('hermes')
+
+        expect(storage.getPendingDeletedSessionIds()).not.toContain('sess-hidden')
+    })
+
+    it('session profile mapping persists across enqueue (until drain succeeds)', async () => {
+        storage.saveSessionProfile('sess-1', 'room-1', 'agent-1', 'hermes')
+        storage.enqueuePendingSessionDelete('sess-1', 'hermes')
+
+        // Re-enqueue (e.g., user tries again) should be idempotent
+        storage.enqueuePendingSessionDelete('sess-1', 'hermes')
+
+        expect(storage.getSessionProfile('sess-1')).not.toBeNull()
+        expect(storage.listPendingSessionDeletes('hermes')).toHaveLength(1)
+    })
+})
diff --git a/tests/server/sessions-db.test.ts b/tests/server/sessions-db.test.ts
index 504b117..d53531f 100644
--- a/tests/server/sessions-db.test.ts
+++ b/tests/server/sessions-db.test.ts
@@ -6,8 +6,8 @@ const contentAllMock = vi.fn()
 const likeAllMock = vi.fn()
 const prepareMock = vi.fn((sql: string) => {
   if (sql.includes('messages_fts MATCH')) return ({ all: contentAllMock })
-  if (sql.includes('m.content LIKE ?')) return ({ all: likeAllMock })
-  if (sql.includes("LOWER(COALESCE(base.title, '')) LIKE ?")) return ({ all: titleAllMock })
+  if (sql.includes('JOIN messages m') && sql.includes('LIKE')) return ({ all: likeAllMock })
+  if (sql.includes('base.title') && sql.includes('LIKE')) return ({ all: titleAllMock })
   return ({ all: allMock })
 })
 const closeMock = vi.fn()
@@ -231,25 +231,25 @@ describe('session DB summaries', () => {
     expect(rows[1].snippet).toContain('docker')
   })
 
-  it('falls back to LIKE search when messages_fts is missing for numeric queries', async () => {
+  it('falls back to literal content search for punctuation-only queries instead of unsafe FTS', async () => {
     titleAllMock.mockReturnValue([])
     contentAllMock.mockImplementation(() => {
-      throw new Error('no such table: messages_fts')
+      throw new Error('fts5: syntax error near "."')
     })
     likeAllMock.mockReturnValue([
       {
-        id: 'numeric-1',
+        id: 'dot-1',
         source: 'cli',
         user_id: '',
         model: 'openai/gpt-5.4',
         title: '',
-        started_at: 1710002800,
+        started_at: 1710004000,
         ended_at: null,
         end_reason: '',
         message_count: 1,
         tool_call_count: 0,
-        input_tokens: 2,
-        output_tokens: 3,
+        input_tokens: 1,
+        output_tokens: 1,
         cache_read_tokens: 0,
         cache_write_tokens: 0,
         reasoning_tokens: 0,
@@ -257,42 +257,39 @@ describe('session DB summaries', () => {
         estimated_cost_usd: 0,
         actual_cost_usd: null,
         cost_status: '',
-        preview: 'numeric preview',
-        last_active: 1710002805,
-        matched_message_id: 9,
-        snippet: 'ticket 12345',
+        preview: 'punctuation preview',
+        last_active: 1710004001,
+        matched_message_id: 21,
+        snippet: 'value.with.dot',
         rank: 0,
       },
     ])
 
     const mod = await import('../../packages/server/src/db/hermes/sessions-db')
-    const rows = await mod.searchSessionSummaries('123', undefined, 10)
+    const rows = await mod.searchSessionSummaries('.', undefined, 10)
 
-    expect(likeAllMock).toHaveBeenCalledWith('123', '%123%')
+    expect(contentAllMock).not.toHaveBeenCalled()
+    expect(likeAllMock).toHaveBeenCalled()
     expect(rows).toHaveLength(1)
-    expect(rows[0].id).toBe('numeric-1')
-    expect(rows[0].snippet).toContain('123')
+    expect(rows[0].id).toBe('dot-1')
   })
 
-  it('keeps the source filter when messages_fts is missing for numeric queries', async () => {
+  it('keeps safe dotted queries on the FTS path', async () => {
     titleAllMock.mockReturnValue([])
-    contentAllMock.mockImplementation(() => {
-      throw new Error('no such table: messages_fts')
-    })
-    likeAllMock.mockReturnValue([
+    contentAllMock.mockReturnValue([
       {
-        id: 'numeric-telegram-1',
-        source: 'telegram',
+        id: 'node-1',
+        source: 'cli',
         user_id: '',
         model: 'openai/gpt-5.4',
-        title: '',
-        started_at: 1710002850,
+        title: 'Node.js notes',
+        started_at: 1710004500,
         ended_at: null,
         end_reason: '',
         message_count: 1,
         tool_call_count: 0,
-        input_tokens: 2,
-        output_tokens: 3,
+        input_tokens: 1,
+        output_tokens: 1,
         cache_read_tokens: 0,
         cache_write_tokens: 0,
         reasoning_tokens: 0,
@@ -300,38 +297,38 @@ describe('session DB summaries', () => {
         estimated_cost_usd: 0,
         actual_cost_usd: null,
         cost_status: '',
-        preview: 'telegram numeric preview',
-        last_active: 1710002855,
-        matched_message_id: 12,
-        snippet: 'telegram 123 body',
-        rank: 0,
+        preview: 'dotted preview',
+        last_active: 1710004501,
+        matched_message_id: 22,
+        snippet: '>>>node.js<<< runtime',
+        rank: 0.2,
       },
     ])
 
     const mod = await import('../../packages/server/src/db/hermes/sessions-db')
-    const rows = await mod.searchSessionSummaries('123', 'telegram', 10)
+    const rows = await mod.searchSessionSummaries('node.js', undefined, 10)
 
-    expect(likeAllMock).toHaveBeenCalledWith('telegram', '123', '%123%')
+    expect(contentAllMock).toHaveBeenCalled()
+    expect(likeAllMock).not.toHaveBeenCalled()
     expect(rows).toHaveLength(1)
-    expect(rows[0].source).toBe('telegram')
-    expect(rows[0].id).toBe('numeric-telegram-1')
+    expect(rows[0].id).toBe('node-1')
   })
 
-  it('preserves title matches when messages_fts is missing for numeric queries', async () => {
+  it('keeps explicit wildcard dotted queries on the FTS path with valid syntax', async () => {
     titleAllMock.mockReturnValue([
       {
-        id: 'title-123',
+        id: 'node-wildcard-title-1',
         source: 'cli',
         user_id: '',
         model: 'openai/gpt-5.4',
-        title: 'Issue 123',
-        started_at: 1710002900,
+        title: 'Node.js wildcard notes',
+        started_at: 1710004590,
         ended_at: null,
         end_reason: '',
         message_count: 1,
         tool_call_count: 0,
-        input_tokens: 2,
-        output_tokens: 3,
+        input_tokens: 1,
+        output_tokens: 1,
         cache_read_tokens: 0,
         cache_write_tokens: 0,
         reasoning_tokens: 0,
@@ -339,30 +336,27 @@ describe('session DB summaries', () => {
         estimated_cost_usd: 0,
         actual_cost_usd: null,
         cost_status: '',
-        preview: 'title numeric preview',
-        last_active: 1710002910,
+        preview: 'wildcard title preview',
+        last_active: 1710004595,
         matched_message_id: null,
-        snippet: 'Issue 123',
+        snippet: 'Node.js wildcard notes',
         rank: 0,
       },
     ])
-    contentAllMock.mockImplementation(() => {
-      throw new Error('no such table: messages_fts')
-    })
-    likeAllMock.mockReturnValue([
+    contentAllMock.mockReturnValue([
       {
-        id: 'content-123',
+        id: 'node-wildcard-1',
         source: 'cli',
         user_id: '',
         model: 'openai/gpt-5.4',
-        title: '',
-        started_at: 1710002890,
+        title: 'Node.js wildcard notes',
+        started_at: 1710004600,
         ended_at: null,
         end_reason: '',
         message_count: 1,
         tool_call_count: 0,
-        input_tokens: 2,
-        output_tokens: 3,
+        input_tokens: 1,
+        output_tokens: 1,
         cache_read_tokens: 0,
         cache_write_tokens: 0,
         reasoning_tokens: 0,
@@ -370,26 +364,243 @@ describe('session DB summaries', () => {
         estimated_cost_usd: 0,
         actual_cost_usd: null,
         cost_status: '',
-        preview: 'content numeric preview',
-        last_active: 1710002895,
-        matched_message_id: 10,
-        snippet: 'content 123 body',
+        preview: 'wildcard dotted preview',
+        last_active: 1710004601,
+        matched_message_id: 24,
+        snippet: '>>>node.js<<< runtime',
+        rank: 0.15,
+      },
+    ])
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+    const rows = await mod.searchSessionSummaries('node.js*', undefined, 10)
+
+    expect(titleAllMock).toHaveBeenCalledWith('%node.js%', 10)
+    expect(contentAllMock).toHaveBeenCalledWith('"node.js"*', 40)
+    expect(likeAllMock).not.toHaveBeenCalled()
+    expect(rows).toHaveLength(2)
+    expect(rows[0].id).toBe('node-wildcard-title-1')
+    expect(rows[1].id).toBe('node-wildcard-1')
+  })
+
+  it('keeps quoted wildcard dotted queries on the FTS path with valid syntax', async () => {
+    titleAllMock.mockReturnValue([
+      {
+        id: 'node-quoted-title-1',
+        source: 'cli',
+        user_id: '',
+        model: 'openai/gpt-5.4',
+        title: 'Quoted Node.js wildcard notes',
+        started_at: 1710004640,
+        ended_at: null,
+        end_reason: '',
+        message_count: 1,
+        tool_call_count: 0,
+        input_tokens: 1,
+        output_tokens: 1,
+        cache_read_tokens: 0,
+        cache_write_tokens: 0,
+        reasoning_tokens: 0,
+        billing_provider: '',
+        estimated_cost_usd: 0,
+        actual_cost_usd: null,
+        cost_status: '',
+        preview: 'quoted title preview',
+        last_active: 1710004645,
+        matched_message_id: null,
+        snippet: 'Quoted Node.js wildcard notes',
+        rank: 0,
+      },
+    ])
+    contentAllMock.mockReturnValue([
+      {
+        id: 'node-quoted-wildcard-1',
+        source: 'cli',
+        user_id: '',
+        model: 'openai/gpt-5.4',
+        title: 'Quoted Node.js wildcard notes',
+        started_at: 1710004650,
+        ended_at: null,
+        end_reason: '',
+        message_count: 1,
+        tool_call_count: 0,
+        input_tokens: 1,
+        output_tokens: 1,
+        cache_read_tokens: 0,
+        cache_write_tokens: 0,
+        reasoning_tokens: 0,
+        billing_provider: '',
+        estimated_cost_usd: 0,
+        actual_cost_usd: null,
+        cost_status: '',
+        preview: 'quoted wildcard dotted preview',
+        last_active: 1710004651,
+        matched_message_id: 25,
+        snippet: '>>>node.js<<< runtime',
+        rank: 0.12,
+      },
+    ])
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+    const rows = await mod.searchSessionSummaries('"node.js"*', undefined, 10)
+
+    expect(titleAllMock).toHaveBeenCalledWith('%node.js%', 10)
+    expect(contentAllMock).toHaveBeenCalledWith('"node.js"*', 40)
+    expect(likeAllMock).not.toHaveBeenCalled()
+    expect(rows).toHaveLength(2)
+    expect(rows[0].id).toBe('node-quoted-title-1')
+    expect(rows[1].id).toBe('node-quoted-wildcard-1')
+  })
+
+  it('keeps non-ASCII dotted queries on the safe quoted FTS path', async () => {
+    titleAllMock.mockReturnValue([])
+    contentAllMock.mockReturnValue([
+      {
+        id: 'unicode-dot-1',
+        source: 'cli',
+        user_id: '',
+        model: 'openai/gpt-5.4',
+        title: 'naïve.js note',
+        started_at: 1710004700,
+        ended_at: null,
+        end_reason: '',
+        message_count: 1,
+        tool_call_count: 0,
+        input_tokens: 1,
+        output_tokens: 1,
+        cache_read_tokens: 0,
+        cache_write_tokens: 0,
+        reasoning_tokens: 0,
+        billing_provider: '',
+        estimated_cost_usd: 0,
+        actual_cost_usd: null,
+        cost_status: '',
+        preview: 'unicode dotted preview',
+        last_active: 1710004701,
+        matched_message_id: 23,
+        snippet: 'naïve.js runtime',
         rank: 0,
       },
     ])
 
     const mod = await import('../../packages/server/src/db/hermes/sessions-db')
-    const rows = await mod.searchSessionSummaries('123', undefined, 10)
+    const rows = await mod.searchSessionSummaries('naïve.js', undefined, 10)
 
-    expect(rows).toHaveLength(2)
-    expect(rows[0].id).toBe('title-123')
-    expect(rows[0].matched_message_id).toBeNull()
-    expect(rows[1].id).toBe('content-123')
-    expect(rows[1].matched_message_id).toBe(10)
+    expect(contentAllMock).toHaveBeenCalledWith('"naïve.js"', 40)
+    expect(likeAllMock).not.toHaveBeenCalled()
+    expect(rows).toHaveLength(1)
+    expect(rows[0].id).toBe('unicode-dot-1')
   })
 
-  it('falls back to LIKE search for CJK queries', async () => {
+  it('escapes LIKE wildcards for literal special-character searches', async () => {
     titleAllMock.mockReturnValue([])
+    likeAllMock.mockReturnValue([
+      {
+        id: 'percent-1',
+        source: 'cli',
+        user_id: '',
+        model: 'openai/gpt-5.4',
+        title: '100% reproducible',
+        started_at: 1710005000,
+        ended_at: null,
+        end_reason: '',
+        message_count: 1,
+        tool_call_count: 0,
+        input_tokens: 1,
+        output_tokens: 1,
+        cache_read_tokens: 0,
+        cache_write_tokens: 0,
+        reasoning_tokens: 0,
+        billing_provider: '',
+        estimated_cost_usd: 0,
+        actual_cost_usd: null,
+        cost_status: '',
+        preview: 'literal percent preview',
+        last_active: 1710005001,
+        matched_message_id: 31,
+        snippet: '100% reproducible',
+        rank: 0,
+      },
+    ])
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+    const rows = await mod.searchSessionSummaries('100%', undefined, 10)
+
+    expect(titleAllMock).toHaveBeenCalledWith('%100\\%%', 10)
+    expect(rows).toHaveLength(1)
+    expect(rows[0].id).toBe('percent-1')
+  })
+
+  it('uses literal search for CJK queries even when FTS returns no rows', async () => {
+    titleAllMock.mockReturnValue([])
+    contentAllMock.mockReturnValue([])
+    likeAllMock.mockReturnValue([
+      {
+        id: 'cjk-literal-1',
+        source: 'cli',
+        user_id: '',
+        model: 'openai/gpt-5.4',
+        title: '',
+        started_at: 1710002980,
+        ended_at: null,
+        end_reason: '',
+        message_count: 1,
+        tool_call_count: 0,
+        input_tokens: 2,
+        output_tokens: 3,
+        cache_read_tokens: 0,
+        cache_write_tokens: 0,
+        reasoning_tokens: 0,
+        billing_provider: '',
+        estimated_cost_usd: 0,
+        actual_cost_usd: null,
+        cost_status: '',
+        preview: '中文内容预览',
+        last_active: 1710002985,
+        matched_message_id: 10,
+        snippet: '这里也有记忆断裂',
+        rank: 0,
+      },
+    ])
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+    const rows = await mod.searchSessionSummaries('记忆断裂', undefined, 10)
+
+    expect(contentAllMock).not.toHaveBeenCalled()
+    expect(likeAllMock).toHaveBeenCalledWith('记忆断裂', '%记忆断裂%', 40)
+    expect(rows).toHaveLength(1)
+    expect(rows[0].id).toBe('cjk-literal-1')
+  })
+
+  it('falls back to LIKE search for CJK queries while preserving title matches', async () => {
+    titleAllMock.mockReturnValue([
+      {
+        id: 'cjk-title-1',
+        source: 'cli',
+        user_id: '',
+        model: 'openai/gpt-5.4',
+        title: '记忆断裂标题',
+        started_at: 1710002990,
+        ended_at: null,
+        end_reason: '',
+        message_count: 1,
+        tool_call_count: 0,
+        input_tokens: 2,
+        output_tokens: 2,
+        cache_read_tokens: 0,
+        cache_write_tokens: 0,
+        reasoning_tokens: 0,
+        billing_provider: '',
+        estimated_cost_usd: 0,
+        actual_cost_usd: null,
+        cost_status: '',
+        preview: 'title preview',
+        last_active: 1710002995,
+        matched_message_id: null,
+        snippet: '记忆断裂标题',
+        rank: 0,
+      },
+    ])
     contentAllMock.mockImplementation(() => {
       throw new Error('fts5 tokenizer miss')
     })
@@ -425,12 +636,96 @@ describe('session DB summaries', () => {
     const mod = await import('../../packages/server/src/db/hermes/sessions-db')
     const rows = await mod.searchSessionSummaries('记忆断裂', undefined, 10)
 
-    expect(likeAllMock).toHaveBeenCalledWith('记忆断裂', '%记忆断裂%')
-    expect(rows).toHaveLength(1)
+    expect(likeAllMock).toHaveBeenCalledWith('记忆断裂', '%记忆断裂%', 40)
+    expect(rows).toHaveLength(2)
     expect(rows[0].id).toBe('cjk-1')
+    expect(rows[1].id).toBe('cjk-title-1')
     expect(rows[0].snippet).toContain('记忆断裂')
   })
 
+  it('does not hide real database failures for safe FTS queries', async () => {
+    titleAllMock.mockReturnValue([])
+    contentAllMock.mockImplementation(() => {
+      throw new Error('database malformed')
+    })
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+
+    await expect(mod.searchSessionSummaries('docker', undefined, 10)).rejects.toThrow(
+      'Failed to search sessions: database malformed',
+    )
+    expect(likeAllMock).not.toHaveBeenCalled()
+  })
+
+  it('throws when messages_fts is missing for numeric queries', async () => {
+    titleAllMock.mockReturnValue([])
+    contentAllMock.mockImplementation(() => {
+      throw new Error('no such table: messages_fts')
+    })
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+
+    await expect(mod.searchSessionSummaries('123', undefined, 10)).rejects.toThrow(
+      'Failed to search sessions: no such table: messages_fts',
+    )
+    expect(likeAllMock).not.toHaveBeenCalled()
+  })
+
+  it('throws when messages_fts is missing for numeric queries with source filter', async () => {
+    titleAllMock.mockReturnValue([])
+    contentAllMock.mockImplementation(() => {
+      throw new Error('no such table: messages_fts')
+    })
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+
+    await expect(mod.searchSessionSummaries('123', 'telegram', 10)).rejects.toThrow(
+      'Failed to search sessions: no such table: messages_fts',
+    )
+    expect(likeAllMock).not.toHaveBeenCalled()
+  })
+
+  it('throws when messages_fts is missing for numeric queries even with title matches', async () => {
+    titleAllMock.mockReturnValue([
+      {
+        id: 'title-123',
+        source: 'cli',
+        user_id: '',
+        model: 'openai/gpt-5.4',
+        title: 'Issue 123',
+        started_at: 1710002900,
+        ended_at: null,
+        end_reason: '',
+        message_count: 1,
+        tool_call_count: 0,
+        input_tokens: 2,
+        output_tokens: 3,
+        cache_read_tokens: 0,
+        cache_write_tokens: 0,
+        reasoning_tokens: 0,
+        billing_provider: '',
+        estimated_cost_usd: 0,
+        actual_cost_usd: null,
+        cost_status: '',
+        preview: 'title numeric preview',
+        last_active: 1710002910,
+        matched_message_id: null,
+        snippet: 'Issue 123',
+        rank: 0,
+      },
+    ])
+    contentAllMock.mockImplementation(() => {
+      throw new Error('no such table: messages_fts')
+    })
+
+    const mod = await import('../../packages/server/src/db/hermes/sessions-db')
+
+    await expect(mod.searchSessionSummaries('123', undefined, 10)).rejects.toThrow(
+      'Failed to search sessions: no such table: messages_fts',
+    )
+    expect(likeAllMock).not.toHaveBeenCalled()
+  })
+
   it('does not fall back to LIKE when messages_fts is missing for non-numeric queries', async () => {
     titleAllMock.mockReturnValue([])
     contentAllMock.mockImplementation(() => {
diff --git a/vite.config.ts b/vite.config.ts
index 17b14d8..d6359a4 100644
--- a/vite.config.ts
+++ b/vite.config.ts
@@ -48,6 +48,10 @@ export default defineConfig({
       '/health': createProxyConfig(),
       '/upload': createProxyConfig(),
       '/webhook': createProxyConfig(),
+      '/socket.io': {
+        target: BACKEND,
+        ws: true,
+      },
     },
   },
 })