diff --git a/packages/client/src/components/hermes/chat/ChatPanel.vue b/packages/client/src/components/hermes/chat/ChatPanel.vue
index 3279070f..0c14274a 100644
--- a/packages/client/src/components/hermes/chat/ChatPanel.vue
+++ b/packages/client/src/components/hermes/chat/ChatPanel.vue
@@ -11,7 +11,15 @@ const chatStore = useChatStore()
 const message = useMessage()
 const { t } = useI18n()
 
-const showSessions = ref(true)
+// Initialize synchronously from the media query so first paint is correct.
+// On narrow viewports the session list is an absolute-positioned overlay
+// (z-index 10) on top of the chat area; if we default to `true`, onMounted
+// only flips it to `false` AFTER the first render, causing a visible flash
+// where the session list covers the chat content ("auto-fixes after a
+// moment" — that was the race).
+const showSessions = ref(
+  typeof window === 'undefined' || !window.matchMedia('(max-width: 768px)').matches,
+)
 let mobileQuery: MediaQueryList | null = null
 
 function handleSessionClick(sessionId: string) {
diff --git a/packages/client/src/components/hermes/chat/MessageList.vue b/packages/client/src/components/hermes/chat/MessageList.vue
index b1311f8e..03a177e9 100644
--- a/packages/client/src/components/hermes/chat/MessageList.vue
+++ b/packages/client/src/components/hermes/chat/MessageList.vue
@@ -45,7 +45,7 @@ watch(
   scrollToBottom,
 );
 watch(
-  () => chatStore.isStreaming,
+  () => chatStore.isRunActive,
   (v) => {
     if (v) scrollToBottom();
   },
@@ -61,7 +61,7 @@ watch(currentToolCalls, scrollToBottom);
     </div>
     <MessageItem v-for="msg in displayMessages" :key="msg.id" :message="msg" />
     <Transition name="fade">
-      <div v-if="chatStore.isStreaming" class="streaming-indicator">
+      <div v-if="chatStore.isRunActive" class="streaming-indicator">
         <video
           :src="isDark ? thinkingVideoDark : thinkingVideoLight"
           autoplay
diff --git a/packages/client/src/i18n/locales/en.ts b/packages/client/src/i18n/locales/en.ts
index 6a594441..337fcb61 100644
--- a/packages/client/src/i18n/locales/en.ts
+++ b/packages/client/src/i18n/locales/en.ts
@@ -61,6 +61,8 @@ export default {
     settings: 'Settings',
     connected: 'Connected',
     disconnected: 'Disconnected',
+    collapse: 'Collapse menu',
+    expand: 'Expand menu',
     updateTip: 'Run "hermes-web-ui update" in terminal to update',
     updateVersion: 'Upgrade to v{version}',
     updating: 'Updating...',
diff --git a/packages/client/src/i18n/locales/zh.ts b/packages/client/src/i18n/locales/zh.ts
index dc71fc85..6fa017a6 100644
--- a/packages/client/src/i18n/locales/zh.ts
+++ b/packages/client/src/i18n/locales/zh.ts
@@ -61,6 +61,8 @@ export default {
     settings: '设置',
     connected: '已连接',
     disconnected: '未连接',
+    collapse: '收起菜单',
+    expand: '展开菜单',
     updateTip: '在终端运行 "hermes-web-ui update" 即可更新',
     updateVersion: '升级版本 v{version}',
     updating: '正在更新...',
diff --git a/packages/client/src/stores/hermes/app.ts b/packages/client/src/stores/hermes/app.ts
index b9d76571..3206decc 100644
--- a/packages/client/src/stores/hermes/app.ts
+++ b/packages/client/src/stores/hermes/app.ts
@@ -4,8 +4,12 @@ import { checkHealth, fetchAvailableModels, updateDefaultModel, triggerUpdate, t
 
 const WEB_UI_VERSION = __APP_VERSION__
 
+const SIDEBAR_COLLAPSED_KEY = 'hermes_sidebar_collapsed'
+
 export const useAppStore = defineStore('app', () => {
   const sidebarOpen = ref(false)
+  // Desktop-only collapsed state (icon-rail mode). Persisted to localStorage.
+  const sidebarCollapsed = ref(localStorage.getItem(SIDEBAR_COLLAPSED_KEY) === '1')
 
   const connected = ref(false)
   const serverVersion = ref(WEB_UI_VERSION)
@@ -90,10 +94,21 @@ export const useAppStore = defineStore('app', () => {
     sidebarOpen.value = false
   }
 
+  function toggleSidebarCollapsed() {
+    sidebarCollapsed.value = !sidebarCollapsed.value
+    try {
+      localStorage.setItem(SIDEBAR_COLLAPSED_KEY, sidebarCollapsed.value ? '1' : '0')
+    } catch {
+      // ignore quota errors — fallback to in-memory only
+    }
+  }
+
   return {
     sidebarOpen,
+    sidebarCollapsed,
     toggleSidebar,
     closeSidebar,
+    toggleSidebarCollapsed,
     connected,
     serverVersion,
     latestVersion,
diff --git a/packages/client/src/stores/hermes/chat.ts b/packages/client/src/stores/hermes/chat.ts
index cedc73ed..170d41e6 100644
--- a/packages/client/src/stores/hermes/chat.ts
+++ b/packages/client/src/stores/hermes/chat.ts
@@ -159,23 +159,244 @@ function mapHermesSession(s: SessionSummary): Session {
   }
 }
 
+// Cache keys for stale-while-revalidate loading of sessions / messages.
+// Rendering from cache on boot avoids the multi-round-trip wait the user sees
+// every time they open the page (esp. noticeable on mobile).
+const STORAGE_KEY = 'hermes_active_session'
+const SESSIONS_CACHE_KEY = 'hermes_sessions_cache_v1'
+const MSGS_CACHE_KEY_PREFIX = 'hermes_session_msgs_v1_'
+// tmux-like resume: persist active run info so a refresh/reopen mid-run can
+// pick up the working indicator and poll fetchSession for new progress.
+const IN_FLIGHT_KEY_PREFIX = 'hermes_in_flight_v1_'
+const IN_FLIGHT_TTL_MS = 15 * 60 * 1000 // Give up after 15 minutes
+const POLL_INTERVAL_MS = 2000
+const POLL_STABLE_EXITS = 3 // 3 × 2s = 6s of no change → assume run finished
+
+interface InFlightRun {
+  runId: string
+  startedAt: number
+}
+
+function loadJson<T>(key: string): T | null {
+  try {
+    const raw = localStorage.getItem(key)
+    return raw ? (JSON.parse(raw) as T) : null
+  } catch {
+    return null
+  }
+}
+
+function saveJson(key: string, value: unknown) {
+  try {
+    localStorage.setItem(key, JSON.stringify(value))
+  } catch {
+    // quota exceeded or private mode — ignore, cache is best-effort
+  }
+}
+
+function removeItem(key: string) {
+  try {
+    localStorage.removeItem(key)
+  } catch {
+    // ignore
+  }
+}
+
+// Strip the circular `file: File` reference from attachments before caching —
+// File objects don't serialize and we only need name/type/size/url for display.
+function sanitizeForCache(msgs: Message[]): Message[] {
+  return msgs.map(m => {
+    if (!m.attachments?.length) return m
+    return {
+      ...m,
+      attachments: m.attachments.map(a => ({ id: a.id, name: a.name, type: a.type, size: a.size, url: a.url })),
+    }
+  })
+}
+
 export const useChatStore = defineStore('chat', () => {
-  const STORAGE_KEY = 'hermes_active_session'
   const sessions = ref<Session[]>([])
   const activeSessionId = ref<string | null>(localStorage.getItem(STORAGE_KEY))
   const streamStates = ref<Map<string, AbortController>>(new Map())
   const isStreaming = computed(() => activeSessionId.value != null && streamStates.value.has(activeSessionId.value))
   const isLoadingSessions = ref(false)
   const isLoadingMessages = ref(false)
+  // tmux-like resume state: true when we recovered an in-flight run from
+  // localStorage after a refresh and are polling fetchSession for progress.
+  // UI shows the thinking indicator while this is set.
+  const resumingRuns = ref<Set<string>>(new Set())
+  const isRunActive = computed(() =>
+    isStreaming.value
+    || (activeSessionId.value != null && resumingRuns.value.has(activeSessionId.value))
+  )
+  const pollTimers = new Map<string, ReturnType<typeof setInterval>>()
+  const pollSignatures = new Map<string, { sig: string, stableTicks: number }>()
 
   const activeSession = ref<Session | null>(null)
   const messages = computed<Message[]>(() => activeSession.value?.messages || [])
 
+  // Hydrate from cache synchronously so the UI renders instantly on boot.
+  // Network revalidation happens in loadSessions() below.
+  const cachedSessions = loadJson<Session[]>(SESSIONS_CACHE_KEY)
+  if (cachedSessions?.length) {
+    sessions.value = cachedSessions
+    if (activeSessionId.value) {
+      const cachedActive = cachedSessions.find(s => s.id === activeSessionId.value) || null
+      if (cachedActive) {
+        const cachedMsgs = loadJson<Message[]>(MSGS_CACHE_KEY_PREFIX + activeSessionId.value)
+        if (cachedMsgs) cachedActive.messages = cachedMsgs
+        activeSession.value = cachedActive
+      }
+    }
+  }
+
+  function persistSessionsList() {
+    // Cache lightweight summaries only (messages are cached per-session).
+    saveJson(
+      SESSIONS_CACHE_KEY,
+      sessions.value.map(s => ({ ...s, messages: [] })),
+    )
+  }
+
+  function persistActiveMessages() {
+    const sid = activeSessionId.value
+    if (!sid) return
+    const s = sessions.value.find(sess => sess.id === sid)
+    if (s) saveJson(MSGS_CACHE_KEY_PREFIX + sid, sanitizeForCache(s.messages))
+  }
+
+  function markInFlight(sid: string, runId: string) {
+    saveJson(IN_FLIGHT_KEY_PREFIX + sid, { runId, startedAt: Date.now() } as InFlightRun)
+  }
+
+  function clearInFlight(sid: string) {
+    removeItem(IN_FLIGHT_KEY_PREFIX + sid)
+  }
+
+  function readInFlight(sid: string): InFlightRun | null {
+    const rec = loadJson<InFlightRun>(IN_FLIGHT_KEY_PREFIX + sid)
+    if (!rec) return null
+    if (Date.now() - rec.startedAt > IN_FLIGHT_TTL_MS) {
+      removeItem(IN_FLIGHT_KEY_PREFIX + sid)
+      return null
+    }
+    return rec
+  }
+
+  function stopPolling(sid: string) {
+    const t = pollTimers.get(sid)
+    if (t) {
+      clearInterval(t)
+      pollTimers.delete(sid)
+    }
+    pollSignatures.delete(sid)
+    resumingRuns.value = new Set([...resumingRuns.value].filter(x => x !== sid))
+  }
+
+  // Poll fetchSession while an in-flight run is recovering. Exits when the
+  // server's message signature is stable for POLL_STABLE_EXITS ticks (run
+  // presumed done), TTL elapses, or the user explicitly starts streaming.
+  function startPolling(sid: string) {
+    if (pollTimers.has(sid)) return
+    resumingRuns.value = new Set([...resumingRuns.value, sid])
+    const timer = setInterval(async () => {
+      // If a fresh SSE stream started for this session, polling is redundant.
+      if (streamStates.value.has(sid)) {
+        stopPolling(sid)
+        return
+      }
+      const inFlight = readInFlight(sid)
+      if (!inFlight) {
+        stopPolling(sid)
+        return
+      }
+      try {
+        const detail = await fetchSession(sid)
+        if (!detail) return
+        const mapped = mapHermesMessages(detail.messages || [])
+        const target = sessions.value.find(s => s.id === sid)
+        if (!target) return
+        // Use the same "content-aware" comparison as switchSession: server
+        // is ahead iff it knows about at least as many user turns and its
+        // last assistant text is at least as long as ours.
+        const local = target.messages
+        const localLastAssistant = [...local].reverse().find(m => m.role === 'assistant')
+        const serverLastAssistant = [...mapped].reverse().find(m => m.role === 'assistant')
+        const localAssistantLen = localLastAssistant?.content?.length ?? 0
+        const serverAssistantLen = serverLastAssistant?.content?.length ?? 0
+        const localUsers = local.filter(m => m.role === 'user').length
+        const serverUsers = mapped.filter(m => m.role === 'user').length
+        const serverIsCaughtUp = serverUsers >= localUsers
+        // Same rationale as switchSession: strictly more user turns means
+        // server is ahead (new turn complete). Equal user turns + longer
+        // assistant means server caught up on the current turn.
+        const serverIsAhead =
+          serverUsers > localUsers
+          || (serverUsers === localUsers && serverAssistantLen >= localAssistantLen)
+        if (serverIsAhead) {
+          target.messages = mapped
+          target.inputTokens = detail.input_tokens
+          target.outputTokens = detail.output_tokens
+          if (detail.title && !target.title) target.title = detail.title
+          if (sid === activeSessionId.value) persistActiveMessages()
+        }
+        // Stability detection ONLY matters when the server has at least as
+        // many user turns as we do. Otherwise the server is still catching
+        // up (e.g. the new turn we just sent hasn't been flushed server-side
+        // yet) and a "stable" signature is a false positive — the stability
+        // is the server NOT having our latest turn, not the run being done.
+        if (!serverIsCaughtUp) {
+          pollSignatures.delete(sid)
+        } else {
+          const last = mapped[mapped.length - 1]
+          const sig = `${mapped.length}|${last?.content?.slice(-40) || ''}|${last?.toolStatus || ''}`
+          const prev = pollSignatures.get(sid)
+          if (prev && prev.sig === sig) {
+            prev.stableTicks += 1
+            if (prev.stableTicks >= POLL_STABLE_EXITS) {
+              // Run is done on the server. Force-apply server view even if
+              // our "don't retreat" guard above skipped it — the server is
+              // now the authoritative source of truth.
+              target.messages = mapped
+              target.inputTokens = detail.input_tokens
+              target.outputTokens = detail.output_tokens
+              if (detail.title) target.title = detail.title
+              if (sid === activeSessionId.value) persistActiveMessages()
+              clearInFlight(sid)
+              stopPolling(sid)
+            }
+          } else {
+            pollSignatures.set(sid, { sig, stableTicks: 0 })
+          }
+        }
+      } catch {
+        // transient network error — ignore, next tick tries again
+      }
+    }, POLL_INTERVAL_MS)
+    pollTimers.set(sid, timer)
+  }
+
   async function loadSessions() {
     isLoadingSessions.value = true
     try {
       const list = await fetchSessions()
-      sessions.value = list.map(mapHermesSession)
+      const fresh = list.map(mapHermesSession)
+      const freshIds = new Set(fresh.map(s => s.id))
+      // Preserve already-loaded messages for sessions that are still present,
+      // so we don't blow away the active session's messages on refresh.
+      const msgsByIdBefore = new Map(sessions.value.map(s => [s.id, s.messages]))
+      for (const s of fresh) {
+        const prev = msgsByIdBefore.get(s.id)
+        if (prev && prev.length) s.messages = prev
+      }
+      // Preserve local-only sessions the server hasn't seen yet — e.g. a 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.value = [...localOnly, ...fresh]
+      persistSessionsList()
+
       // Restore last active session, fallback to most recent
       const savedId = activeSessionId.value
       const targetId = savedId && sessions.value.some(s => s.id === savedId)
@@ -191,6 +412,30 @@ export const useChatStore = defineStore('chat', () => {
     }
   }
 
+  // Re-pull active session from server and overwrite local messages. Used on
+  // SSE drop and on tab-visible events — mobile browsers kill EventSource
+  // while backgrounded, but the backend run usually completes anyway.
+  async function refreshActiveSession(): Promise<boolean> {
+    const sid = activeSessionId.value
+    if (!sid) return false
+    try {
+      const detail = await fetchSession(sid)
+      if (!detail) return false
+      const target = sessions.value.find(s => s.id === sid)
+      if (!target) return false
+      const mapped = mapHermesMessages(detail.messages || [])
+      target.messages = mapped
+      target.inputTokens = detail.input_tokens
+      target.outputTokens = detail.output_tokens
+      if (detail.title) target.title = detail.title
+      persistActiveMessages()
+      return true
+    } catch (err) {
+      console.error('Failed to refresh active session:', err)
+      return false
+    }
+  }
+
 
   function createSession(): Session {
     const session: Session = {
@@ -202,6 +447,9 @@ export const useChatStore = defineStore('chat', () => {
       updatedAt: Date.now(),
     }
     sessions.value.unshift(session)
+    // Persist immediately so a refresh before run.completed can still find
+    // this session in the cache.
+    persistSessionsList()
     return session
   }
 
@@ -210,32 +458,81 @@ export const useChatStore = defineStore('chat', () => {
     localStorage.setItem(STORAGE_KEY, sessionId)
     activeSession.value = sessions.value.find(s => s.id === sessionId) || null
 
-    // If session has no messages loaded, fetch from API
-    if (activeSession.value && activeSession.value.messages.length === 0) {
-      isLoadingMessages.value = true
-      try {
-        const detail = await fetchSession(sessionId)
-        if (detail && detail.messages) {
-          const mapped = mapHermesMessages(detail.messages)
+    if (!activeSession.value) return
+
+    // Hydrate messages from localStorage cache first (instant render), then
+    // revalidate from server in the background. If no cache exists, show the
+    // loading state while we fetch.
+    const hasLocalMessages = activeSession.value.messages.length > 0
+    if (!hasLocalMessages) {
+      const cachedMsgs = loadJson<Message[]>(MSGS_CACHE_KEY_PREFIX + sessionId)
+      if (cachedMsgs?.length) {
+        activeSession.value.messages = cachedMsgs
+      }
+    }
+
+    const needsBlockingLoad = activeSession.value.messages.length === 0
+    if (needsBlockingLoad) isLoadingMessages.value = true
+
+    try {
+      const detail = await fetchSession(sessionId)
+      if (detail && detail.messages) {
+        const mapped = mapHermesMessages(detail.messages)
+        // Pick whichever view has more information. Simple length comparison
+        // is wrong because mapHermesMessages folds tool_call-only assistant
+        // msgs and matches them with tool-result msgs — so post-fold `mapped`
+        // can be SHORTER than the raw SSE-built local array even when the
+        // server is strictly ahead. Instead, compare the last assistant
+        // message content: if the server's is at least as long, the server
+        // is up-to-date (and has the final complete text); otherwise keep
+        // local (in-flight window where server hasn't flushed the new turn).
+        const local = activeSession.value.messages
+        const localLastAssistant = [...local].reverse().find(m => m.role === 'assistant')
+        const serverLastAssistant = [...mapped].reverse().find(m => m.role === 'assistant')
+        const localAssistantLen = localLastAssistant?.content?.length ?? 0
+        const serverAssistantLen = serverLastAssistant?.content?.length ?? 0
+        const localUsers = local.filter(m => m.role === 'user').length
+        const serverUsers = mapped.filter(m => m.role === 'user').length
+        // Trust server when:
+        //   - it has STRICTLY MORE user turns than we do (new turn landed),
+        //     OR
+        //   - same user-turn count AND server's last assistant is at least
+        //     as long as ours (same turn, server caught up or further)
+        // Otherwise keep local (protects against the server-not-yet-flushed
+        // race during in-flight runs). Length comparison alone is wrong
+        // across different turns because each turn's last assistant is
+        // unrelated to the previous turn's.
+        const serverIsAhead =
+          serverUsers > localUsers
+          || (serverUsers === localUsers && serverAssistantLen >= localAssistantLen)
+        if (serverIsAhead) {
           activeSession.value.messages = mapped
-          activeSession.value.inputTokens = detail.input_tokens
-          activeSession.value.outputTokens = detail.output_tokens
-          // Update title: use Hermes title, or fallback to first user message
-          if (detail.title) {
-            activeSession.value.title = detail.title
-          } else {
-            const firstUser = mapped.find(m => m.role === 'user')
-            if (firstUser) {
-              const t = firstUser.content.slice(0, 40)
-              activeSession.value.title = t + (firstUser.content.length > 40 ? '...' : '')
-            }
+        }
+        activeSession.value.inputTokens = detail.input_tokens
+        activeSession.value.outputTokens = detail.output_tokens
+        // Update title: use Hermes title, or fallback to first user message
+        if (detail.title) {
+          activeSession.value.title = detail.title
+        } else if (!activeSession.value.title) {
+          const firstUser = (activeSession.value.messages).find(m => m.role === 'user')
+          if (firstUser) {
+            const t = firstUser.content.slice(0, 40)
+            activeSession.value.title = t + (firstUser.content.length > 40 ? '...' : '')
           }
         }
-      } catch (err) {
-        console.error('Failed to load session messages:', err)
-      } finally {
-        isLoadingMessages.value = false
+        persistActiveMessages()
       }
+    } catch (err) {
+      console.error('Failed to load session messages:', err)
+    } finally {
+      isLoadingMessages.value = false
+    }
+
+    // tmux-like resume: if this session has a recent in-flight run and we're
+    // not currently streaming, start polling fetchSession to pick up progress
+    // that happened while we were gone. Exits automatically on stability.
+    if (readInFlight(sessionId) && !streamStates.value.has(sessionId)) {
+      startPolling(sessionId)
     }
   }
 
@@ -262,6 +559,8 @@ export const useChatStore = defineStore('chat', () => {
   async function deleteSession(sessionId: string) {
     await deleteSessionApi(sessionId)
     sessions.value = sessions.value.filter(s => s.id !== sessionId)
+    removeItem(MSGS_CACHE_KEY_PREFIX + sessionId)
+    persistSessionsList()
     if (activeSessionId.value === sessionId) {
       if (sessions.value.length > 0) {
         await switchSession(sessions.value[0].id)
@@ -326,6 +625,13 @@ export const useChatStore = defineStore('chat', () => {
     }
     addMessage(sid, userMsg)
     updateSessionTitle(sid)
+    // Persist immediately so a refresh before the first SSE event (e.g. the
+    // user closes the tab right after sending) still has the user's message
+    // and session title in the cache.
+    if (sid === activeSessionId.value) {
+      persistActiveMessages()
+      persistSessionsList()
+    }
 
     try {
       // Build conversation history from past messages
@@ -362,9 +668,33 @@ export const useChatStore = defineStore('chat', () => {
         return
       }
 
+      // tmux-like resume: persist run_id so refresh/reopen can pick up the
+      // working indicator and poll for progress.
+      markInFlight(sid, runId)
+      // If we were already polling (e.g. user re-sent while resume was still
+      // polling an earlier run), cancel that polling — the new SSE stream is
+      // the authoritative live source.
+      stopPolling(sid)
+
       // Helper to clean up this session's stream state
       const cleanup = () => {
         streamStates.value.delete(sid)
+        if (persistTimer) {
+          clearTimeout(persistTimer)
+          persistTimer = null
+        }
+      }
+
+      // Throttle in-flight cache writes so a refresh mid-stream still shows
+      // the partial reply. 800ms keeps quota pressure low while guaranteeing
+      // at most ~1s of unsaved delta on reload.
+      let persistTimer: ReturnType<typeof setTimeout> | null = null
+      const schedulePersist = () => {
+        if (sid !== activeSessionId.value || persistTimer) return
+        persistTimer = setTimeout(() => {
+          persistTimer = null
+          persistActiveMessages()
+        }, 800)
       }
 
       // Listen to SSE events — all closures capture `sid`
@@ -390,6 +720,7 @@ export const useChatStore = defineStore('chat', () => {
                   isStreaming: true,
                 })
               }
+              schedulePersist()
               break
             }
 
@@ -408,6 +739,7 @@ export const useChatStore = defineStore('chat', () => {
                 toolPreview: evt.preview,
                 toolStatus: 'running',
               })
+              schedulePersist()
               break
             }
 
@@ -420,6 +752,7 @@ export const useChatStore = defineStore('chat', () => {
                 const last = toolMsgs[toolMsgs.length - 1]
                 updateMessage(sid, last.id, { toolStatus: 'done' })
               }
+              schedulePersist()
               break
             }
 
@@ -431,6 +764,15 @@ export const useChatStore = defineStore('chat', () => {
               }
               cleanup()
               updateSessionTitle(sid)
+              // IMPORTANT ordering: persist the final cache BEFORE clearing
+              // the in-flight marker. If the browser is reloading right now
+              // and kills us between the two localStorage writes, we want
+              // the next page load to still see in-flight === true (so
+              // polling kicks in and recovers) rather than the other way
+              // around (cleared in-flight + stale streaming cache = UI stuck).
+              if (sid === activeSessionId.value) persistActiveMessages()
+              clearInFlight(sid)
+              stopPolling(sid)
               break
             }
 
@@ -457,6 +799,9 @@ export const useChatStore = defineStore('chat', () => {
                 }
               })
               cleanup()
+              if (sid === activeSessionId.value) persistActiveMessages()
+              clearInFlight(sid)
+              stopPolling(sid)
               break
             }
           }
@@ -472,24 +817,36 @@ export const useChatStore = defineStore('chat', () => {
           updateSessionTitle(sid)
         },
         // onError
+        // Mobile browsers drop EventSource when the tab backgrounds / screen
+        // locks / network flips. The backend run usually completes anyway, so
+        // rather than injecting a stale "SSE connection error" bubble we mark
+        // streaming as done and silently re-sync from the server, which has
+        // the real final answer. If the server fetch itself fails, we leave
+        // whatever text we already streamed in place — no visible error.
         (err) => {
+          console.warn('SSE connection dropped, resyncing from server:', err.message)
           const msgs = getSessionMsgs(sid)
           const last = msgs[msgs.length - 1]
           if (last?.isStreaming) {
-            updateMessage(sid, last.id, {
-              isStreaming: false,
-              content: `Error: ${err.message}`,
-              role: 'system',
-            })
-          } else {
-            addMessage(sid, {
-              id: uid(),
-              role: 'system',
-              content: `Error: ${err.message}`,
-              timestamp: Date.now(),
-            })
+            updateMessage(sid, last.id, { isStreaming: false })
           }
+          // Any tool messages still marked 'running' will be replaced by the
+          // server's view after refresh; clear their spinner state now.
+          msgs.forEach((m, i) => {
+            if (m.role === 'tool' && m.toolStatus === 'running') {
+              msgs[i] = { ...m, toolStatus: 'done' }
+            }
+          })
           cleanup()
+          if (sid === activeSessionId.value) {
+            void refreshActiveSession()
+          }
+          // The run might still be going on the server side (SSE drop doesn't
+          // abort it). If we still have an in-flight record, fall back to
+          // polling fetchSession to keep the user updated.
+          if (readInFlight(sid)) {
+            startPolling(sid)
+          }
         },
       )
 
@@ -516,18 +873,48 @@ export const useChatStore = defineStore('chat', () => {
         updateMessage(sid, lastMsg.id, { isStreaming: false })
       }
       streamStates.value.delete(sid)
+      clearInFlight(sid)
+      stopPolling(sid)
     }
   }
 
-  // Load sessions on init
+  // Load sessions on init (cache has already hydrated the UI above).
   loadSessions()
 
+  // tmux-like resume on boot: if the last active session has a persisted
+  // in-flight run that's still fresh, show the working indicator immediately
+  // and start polling the server. loadSessions() above will call
+  // switchSession which also triggers this path, but doing it synchronously
+  // here means the UI shows "working" from the very first frame even while
+  // loadSessions is still in flight.
+  if (activeSessionId.value && readInFlight(activeSessionId.value)) {
+    startPolling(activeSessionId.value)
+  }
+
+  // When the tab returns to the foreground, re-sync the active session from
+  // the server. Mobile browsers suspend tabs aggressively, and any in-flight
+  // run that completed while we were backgrounded won't have reached the
+  // in-memory state otherwise.
+  if (typeof document !== 'undefined') {
+    document.addEventListener('visibilitychange', () => {
+      if (document.visibilityState === 'visible' && activeSessionId.value && !isStreaming.value) {
+        void refreshActiveSession()
+        // Resume polling too in case we put a run in-flight before going to
+        // background and the SSE got killed.
+        if (readInFlight(activeSessionId.value)) {
+          startPolling(activeSessionId.value)
+        }
+      }
+    })
+  }
+
   return {
     sessions,
     activeSessionId,
     activeSession,
     messages,
     isStreaming,
+    isRunActive,
     isLoadingSessions,
     isLoadingMessages,
     newChat,
@@ -537,5 +924,6 @@ export const useChatStore = defineStore('chat', () => {
     sendMessage,
     stopStreaming,
     loadSessions,
+    refreshActiveSession,
   }
 })
diff --git a/tests/client/app-store.test.ts b/tests/client/app-store.test.ts
new file mode 100644
index 00000000..02856c57
--- /dev/null
+++ b/tests/client/app-store.test.ts
@@ -0,0 +1,36 @@
+// @vitest-environment jsdom
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+import { createPinia, setActivePinia } from 'pinia'
+
+const mockSystemApi = vi.hoisted(() => ({
+  checkHealth: vi.fn(),
+  fetchAvailableModels: vi.fn(),
+  updateDefaultModel: vi.fn(),
+  triggerUpdate: vi.fn(),
+}))
+
+vi.mock('@/api/hermes/system', () => mockSystemApi)
+
+import { useAppStore } from '@/stores/hermes/app'
+
+describe('App Store', () => {
+  beforeEach(() => {
+    setActivePinia(createPinia())
+    vi.clearAllMocks()
+    window.localStorage.clear()
+  })
+
+  it('persists desktop sidebar collapsed state to localStorage', () => {
+    const store = useAppStore()
+
+    expect(store.sidebarCollapsed).toBe(false)
+
+    store.toggleSidebarCollapsed()
+    expect(store.sidebarCollapsed).toBe(true)
+    expect(window.localStorage.getItem('hermes_sidebar_collapsed')).toBe('1')
+
+    store.toggleSidebarCollapsed()
+    expect(store.sidebarCollapsed).toBe(false)
+    expect(window.localStorage.getItem('hermes_sidebar_collapsed')).toBe('0')
+  })
+})
diff --git a/tests/client/chat-store.test.ts b/tests/client/chat-store.test.ts
new file mode 100644
index 00000000..6bc3c46c
--- /dev/null
+++ b/tests/client/chat-store.test.ts
@@ -0,0 +1,226 @@
+// @vitest-environment jsdom
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+import { createPinia, setActivePinia } from 'pinia'
+
+const mockChatApi = vi.hoisted(() => ({
+  startRun: vi.fn(),
+  streamRunEvents: vi.fn(),
+}))
+
+const mockSessionsApi = vi.hoisted(() => ({
+  fetchSessions: vi.fn(),
+  fetchSession: vi.fn(),
+  deleteSession: vi.fn(),
+  renameSession: vi.fn(),
+}))
+
+vi.mock('@/api/hermes/chat', () => mockChatApi)
+vi.mock('@/api/hermes/sessions', () => mockSessionsApi)
+
+import { useChatStore } from '@/stores/hermes/chat'
+
+function makeSummary(id: string, title = 'Session') {
+  return {
+    id,
+    source: 'api_server',
+    model: 'gpt-4o',
+    title,
+    started_at: 1710000000,
+    ended_at: 1710000001,
+    message_count: 1,
+    tool_call_count: 0,
+    input_tokens: 10,
+    output_tokens: 20,
+    cache_read_tokens: 0,
+    cache_write_tokens: 0,
+    reasoning_tokens: 0,
+    billing_provider: 'openai',
+    estimated_cost_usd: 0,
+    actual_cost_usd: 0,
+    cost_status: 'estimated',
+  }
+}
+
+function makeDetail(id: string, messages: Array<Record<string, any>>) {
+  return {
+    ...makeSummary(id),
+    messages,
+  }
+}
+
+async function flushPromises() {
+  await Promise.resolve()
+  await Promise.resolve()
+}
+
+describe('Chat Store', () => {
+  beforeEach(() => {
+    setActivePinia(createPinia())
+    vi.clearAllMocks()
+    vi.useRealTimers()
+    window.localStorage.clear()
+    mockSessionsApi.fetchSessions.mockResolvedValue([])
+    mockSessionsApi.fetchSession.mockResolvedValue(null)
+    mockSessionsApi.deleteSession.mockResolvedValue(true)
+    mockSessionsApi.renameSession.mockResolvedValue(true)
+    mockChatApi.startRun.mockResolvedValue({ run_id: 'run-1', status: 'queued' })
+    mockChatApi.streamRunEvents.mockImplementation(() => ({
+      abort: vi.fn(),
+    }))
+  })
+
+  it('hydrates cached active session immediately and preserves local-only sessions after refresh', async () => {
+    const cachedSession = {
+      id: 'local-1',
+      title: 'Local Draft',
+      source: 'api_server',
+      messages: [],
+      createdAt: 1,
+      updatedAt: 1,
+    }
+    const cachedMessages = [
+      { id: 'm1', role: 'user', content: 'draft', timestamp: 1 },
+    ]
+
+    window.localStorage.setItem('hermes_active_session', 'local-1')
+    window.localStorage.setItem('hermes_sessions_cache_v1', JSON.stringify([cachedSession]))
+    window.localStorage.setItem('hermes_session_msgs_v1_local-1', JSON.stringify(cachedMessages))
+
+    mockSessionsApi.fetchSessions.mockResolvedValue([makeSummary('remote-1', 'Remote Session')])
+    mockSessionsApi.fetchSession.mockResolvedValue(null)
+
+    const store = useChatStore()
+
+    expect(store.activeSessionId).toBe('local-1')
+    expect(store.messages.map(m => m.content)).toEqual(['draft'])
+
+    await flushPromises()
+
+    expect(store.sessions.map(s => s.id)).toEqual(['local-1', 'remote-1'])
+    expect(store.activeSession?.id).toBe('local-1')
+    expect(store.messages.map(m => m.content)).toEqual(['draft'])
+  })
+
+  it('persists the user message immediately before any SSE delta arrives', async () => {
+    const store = useChatStore()
+
+    await flushPromises()
+    await store.sendMessage('hello world')
+
+    const sid = store.activeSessionId
+    expect(sid).toBeTruthy()
+    expect(window.localStorage.getItem('hermes_active_session')).toBe(sid)
+
+    const cachedMessages = JSON.parse(
+      window.localStorage.getItem(`hermes_session_msgs_v1_${sid}`) || '[]',
+    )
+    expect(cachedMessages).toEqual(
+      expect.arrayContaining([
+        expect.objectContaining({
+          role: 'user',
+          content: 'hello world',
+        }),
+      ]),
+    )
+  })
+
+  it('silently refreshes from server on SSE error instead of appending a fake error bubble', async () => {
+    vi.useFakeTimers()
+
+    window.localStorage.setItem('hermes_active_session', 'sess-1')
+    window.localStorage.setItem(
+      'hermes_sessions_cache_v1',
+      JSON.stringify([
+        {
+          id: 'sess-1',
+          title: 'Recovered Chat',
+          source: 'api_server',
+          messages: [],
+          createdAt: 1,
+          updatedAt: 1,
+        },
+      ]),
+    )
+    window.localStorage.setItem(
+      'hermes_session_msgs_v1_sess-1',
+      JSON.stringify([
+        { id: 'old-user', role: 'user', content: 'old prompt', timestamp: 1 },
+      ]),
+    )
+
+    mockSessionsApi.fetchSessions.mockResolvedValue([makeSummary('sess-1', 'Recovered Chat')])
+
+    let fetchSessionCalls = 0
+    mockSessionsApi.fetchSession.mockImplementation(async () => {
+      fetchSessionCalls += 1
+      if (fetchSessionCalls === 1) return null
+      return makeDetail('sess-1', [
+        {
+          id: 1,
+          session_id: 'sess-1',
+          role: 'user',
+          content: 'old prompt',
+          tool_call_id: null,
+          tool_calls: null,
+          tool_name: null,
+          timestamp: 1710000000,
+          token_count: null,
+          finish_reason: null,
+          reasoning: null,
+        },
+        {
+          id: 2,
+          session_id: 'sess-1',
+          role: 'user',
+          content: 'check this',
+          tool_call_id: null,
+          tool_calls: null,
+          tool_name: null,
+          timestamp: 1710000001,
+          token_count: null,
+          finish_reason: null,
+          reasoning: null,
+        },
+        {
+          id: 3,
+          session_id: 'sess-1',
+          role: 'assistant',
+          content: 'final answer',
+          tool_call_id: null,
+          tool_calls: null,
+          tool_name: null,
+          timestamp: 1710000002,
+          token_count: null,
+          finish_reason: 'stop',
+          reasoning: null,
+        },
+      ])
+    })
+
+    mockChatApi.streamRunEvents.mockImplementation((
+      _runId: string,
+      _onEvent: (event: unknown) => void,
+      _onDone: () => void,
+      onError: (err: Error) => void,
+    ) => {
+      setTimeout(() => {
+        onError(new Error('SSE connection error'))
+      }, 0)
+      return { abort: vi.fn() }
+    })
+
+    const store = useChatStore()
+    await flushPromises()
+    await store.sendMessage('check this')
+    await vi.advanceTimersByTimeAsync(0)
+    await flushPromises()
+
+    await vi.advanceTimersByTimeAsync(9000)
+    await flushPromises()
+
+    expect(store.messages.some(m => m.role === 'system' && m.content.includes('SSE connection error'))).toBe(false)
+    expect(store.messages.some(m => m.role === 'assistant' && m.content === 'final answer')).toBe(true)
+    expect(store.isRunActive).toBe(false)
+    expect(window.localStorage.getItem('hermes_in_flight_v1_sess-1')).toBeNull()
+  })
+})
diff --git a/tests/setup.ts b/tests/setup.ts
index 90ab99c3..3915cba0 100644
--- a/tests/setup.ts
+++ b/tests/setup.ts
@@ -1,5 +1,7 @@
 import { vi } from 'vitest'
 
+// Vite injects this at build time; unit tests need a stable fallback.
+;(globalThis as any).__APP_VERSION__ = 'test'
 // Client-only setup (window/localStorage only exist in jsdom)
 if (typeof window !== 'undefined') {
   // Mock window.matchMedia