fix(desktop): keep composer usable during reconnect (#45488)

* feat(cli): add --safe-mode troubleshooting flag

Inspired by Claude Code v2.1.169 (June 2026): run Hermes with all
customizations disabled to isolate setup problems from product bugs.

--safe-mode implies --ignore-user-config and --ignore-rules, and
additionally skips plugin discovery (hermes_cli/plugins.py) and MCP
server loading (tools/mcp_tool.py) via the internal HERMES_SAFE_MODE
env bridge.

* fix(desktop): keep composer usable during reconnect

apps/desktop/src/app/chat/composer/enter-submit-dom-race.test.tsx

@@ -24,6 +24,7 @@ afterEach(cleanup)
 // state stays stale while the DOM already holds the text.
 function Harness({
   busy = false,
+  disabled = false,
   queued = [],
   onSubmit,
   onQueue,
@@ -31,6 +32,7 @@ function Harness({
   onDrain
 }: {
   busy?: boolean
+  disabled?: boolean
   queued?: readonly string[]
   onSubmit: (text: string) => void
   onQueue: (text: string) => void
@@ -52,6 +54,10 @@ function Harness({
   }
 
   const submitDraft = () => {
+    if (disabled) {
+      return
+    }
+
     const editor = editorRef.current
     if (editor) {
       const domText = composerPlainText(editor)
@@ -84,6 +90,10 @@ function Harness({
       const editorText = editorRef.current ? composerPlainText(editorRef.current) : draftRef.current
       const hasLivePayload = editorText.trim().length > 0 || attachments.length > 0
 
+      if (disabled) {
+        return
+      }
+
       if (!busy && !hasLivePayload && queued.length > 0) {
         onDrain()
 
@@ -186,4 +196,23 @@ describe('composer Enter submit — live DOM vs stale composer state (#39630)',
     expect(onDrain).toHaveBeenCalledTimes(1)
     expect(onSubmit).not.toHaveBeenCalled()
   })
+
+  it('keeps reconnect drafts editable but blocks Enter submit until the gateway returns', async () => {
+    const onSubmit = vi.fn()
+    const onDrain = vi.fn()
+    const { getByTestId } = render(
+      <Harness disabled onCancel={vi.fn()} onDrain={onDrain} onQueue={vi.fn()} onSubmit={onSubmit} queued={['queued-1']} />
+    )
+    const editor = getByTestId('editor')
+
+    await act(async () => {
+      editor.textContent = 'draft while reconnecting'
+      fireEvent.input(editor)
+      fireEvent.keyDown(editor, { key: 'Enter' })
+    })
+
+    expect(editor.textContent).toBe('draft while reconnecting')
+    expect(onDrain).not.toHaveBeenCalled()
+    expect(onSubmit).not.toHaveBeenCalled()
+  })
 })

apps/desktop/src/app/chat/composer/index.tsx

@@ -247,6 +247,8 @@ export function ChatBar({
   const gatewayState = useStore($gatewayState)
   const newSessionPlaceholders = t.composer.newSessionPlaceholders
   const followUpPlaceholders = t.composer.followUpPlaceholders
+  const reconnecting = gatewayState === 'closed' || gatewayState === 'error'
+  const inputDisabled = disabled && !reconnecting
 
   // Resting placeholder: a starter for brand-new sessions, a continuation for
   // existing ones. Picked once and only re-rolled when we genuinely move to a
@@ -277,11 +279,13 @@ export function ChatBar({
     setRestingPlaceholder(pickPlaceholder(sessionId ? followUpPlaceholders : newSessionPlaceholders))
   }, [followUpPlaceholders, newSessionPlaceholders, sessionId])
 
-  // When the bar is disabled it's because the gateway isn't open. Distinguish a
-  // cold start ("Starting Hermes...") from a dropped connection we're trying to
-  // restore (e.g. after the Mac slept) so the stuck state reads as recoverable.
+  // When the transport is disabled it's because the gateway isn't open.
+  // Distinguish a cold start ("Starting Hermes...") from a dropped connection
+  // we're trying to restore. During reconnect, keep the textbox editable so a
+  // flaky network doesn't block drafting; only submit/backend actions stay
+  // disabled until the gateway is open again.
   const placeholder = disabled
-    ? gatewayState === 'closed' || gatewayState === 'error'
+    ? reconnecting
       ? t.composer.placeholderReconnecting
       : t.composer.placeholderStarting
     : restingPlaceholder
@@ -323,13 +327,13 @@ export function ChatBar({
   )
 
   useEffect(() => {
-    if (!disabled) {
+    if (!inputDisabled) {
       focusInput()
     }
-  }, [disabled, focusInput, focusKey, focusRequestId])
+  }, [focusInput, focusKey, focusRequestId, inputDisabled])
 
   useEffect(() => {
-    if (disabled) {
+    if (inputDisabled) {
       return undefined
     }
 
@@ -349,7 +353,7 @@ export function ChatBar({
       offFocus()
       offInsert()
     }
-  }, [appendExternalText, disabled])
+  }, [appendExternalText, inputDisabled])
 
   // Keep draftRef in sync with the assistant-ui composer state for callers
   // that read the latest text outside the React render cycle. We don't push
@@ -934,6 +938,10 @@ export function ChatBar({
       const editorText = editorRef.current ? composerPlainText(editorRef.current) : draftRef.current
       const hasLivePayload = editorText.trim().length > 0 || attachments.length > 0
 
+      if (disabled) {
+        return
+      }
+
       if (!busy && !hasLivePayload && queuedPrompts.length > 0) {
         void drainNextQueued()
 
@@ -1476,6 +1484,10 @@ export function ChatBar({
   }
 
   const submitDraft = () => {
+    if (disabled) {
+      return
+    }
+
     // Source the text from the DOM editor, not React state. The AUI composer
     // state (`draft`) and the derived `hasComposerPayload` lag the DOM by a
     // render, so on fast typing or IME composition the final keystroke(s) may
@@ -1656,6 +1668,7 @@ export function ChatBar({
   const input = (
     <div className={cn('relative', stacked ? 'w-full' : 'min-w-(--composer-input-inline-min-width) flex-1')}>
       <div
+        aria-disabled={inputDisabled ? true : undefined}
         aria-label={t.composer.message}
         autoCapitalize="off"
         autoCorrect="off"
@@ -1666,7 +1679,7 @@ export function ChatBar({
           stacked && 'pl-3',
           stacked ? 'w-full' : 'min-w-(--composer-input-inline-min-width) flex-1'
         )}
-        contentEditable={!disabled}
+        contentEditable={!inputDisabled}
         data-placeholder={placeholder}
         data-slot={RICH_INPUT_SLOT}
         onBlur={() => window.setTimeout(closeTrigger, 80)}

apps/desktop/src/components/gateway-connecting-overlay.test.tsx

@@ -3,23 +3,23 @@ import { afterEach, beforeEach, describe, expect, it } from 'vitest'
 
 import { $desktopBoot } from '@/store/boot'
 import { $desktopOnboarding } from '@/store/onboarding'
-import { $gatewayState, setGatewayState } from '@/store/session'
+import { setGatewayState } from '@/store/session'
 
 import { BootFailureOverlay } from './boot-failure-overlay'
 import { GatewayConnectingOverlay } from './gateway-connecting-overlay'
 
 // Repro for the "remote gateway → stuck on CONNECTING, no way to settings"
-// report. The connecting overlay (z-1200, full-screen, pointer-events on) is
-// shown whenever `gatewayState !== 'open' && !boot.error`. The ONLY escape
+// report. The connecting overlay (z-1200, full-screen, pointer-events on) used
+// to be shown whenever `gatewayState !== 'open' && !boot.error`. The ONLY escape
 // hatch — BootFailureOverlay, which has "Use local gateway" / "Sign in" /
 // "Retry" — only renders when `boot.error` is set.
 //
 // useGatewayBoot only calls failDesktopBoot() (which sets boot.error) when the
 // INITIAL boot() throws. After the first successful connect (bootCompleted),
 // any later socket drop goes through scheduleReconnect(), which loops FOREVER
-// against the dead remote and never sets boot.error. So gatewayState sits at
-// 'closed'/'error' with boot.error null → CONNECTING forever, recovery overlay
-// never appears, settings unreachable.
+// against the dead remote. So gatewayState sits at 'closed'/'error' with
+// boot.error null. The fix keeps the initial-boot overlay out of post-boot
+// reconnects, leaving chat/settings usable while the reconnect loop runs.
 
 function resetStores() {
   setGatewayState('idle')
@@ -75,7 +75,7 @@ describe('connecting overlay vs recovery surface', () => {
     expect(isConnectingShown()).toBe(false)
   })
 
-  it('REPRO: remote socket drops AFTER a successful boot → stuck on CONNECTING, no recovery, no settings', () => {
+  it('post-boot socket drops do not re-cover the app with the initial CONNECTING overlay', () => {
     // 1. Initial boot succeeded: gateway opened, boot completed (no error).
     setGatewayState('open')
     const { rerender } = render(
@@ -97,14 +97,14 @@ describe('connecting overlay vs recovery surface', () => {
       </>
     )
 
-    // The connecting overlay reappears and latches...
-    expect(isConnectingShown()).toBe(true)
-    // ...with NO recovery surface, because boot.error was never set.
+    // The initial-boot connecting overlay stays out of the way, so settings and
+    // the composer remain reachable during the reconnect loop.
+    expect(isConnectingShown()).toBe(false)
     expect(isRecoveryShown()).toBe(false)
 
-    // 3. Reconnect loops forever against the dead remote: gatewayState bounces
-    //    closed → error → closed, boot.error never gets set. The user is
-    //    pinned on CONNECTING with no path to Settings indefinitely.
+    // 3. Reconnect loops against the dead remote: gatewayState bounces closed
+    //    → error → closed. Until the escalation path sets boot.error, the app
+    //    remains usable instead of modal-blocked.
     setGatewayState('error')
     rerender(
       <>
@@ -113,7 +113,7 @@ describe('connecting overlay vs recovery surface', () => {
       </>
     )
     expect($desktopBoot.get().error).toBeNull()
-    expect(isConnectingShown()).toBe(true)
+    expect(isConnectingShown()).toBe(false)
     expect(isRecoveryShown()).toBe(false)
   })
 

apps/desktop/src/components/gateway-connecting-overlay.tsx

@@ -52,7 +52,13 @@ export function GatewayConnectingOverlay() {
   const [tail, setTail] = useState(TAIL)
   const [phase, setPhase] = useState<Phase>('live')
 
-  const connecting = gatewayState !== 'open' && !boot.error
+  // The full-screen connecting overlay is for initial boot only. After a
+  // healthy boot, flaky networks / sleep-wake can drop the socket and flip the
+  // gateway state back to closed/error while the app reconnects. Do not cover
+  // the chat then — users should still be able to type drafts, open settings,
+  // and recover instead of staring at a modal CONNECTING screen.
+  const initialBootActive = boot.visible || boot.running || boot.progress < 100
+  const connecting = gatewayState !== 'open' && !boot.error && initialBootActive
   // Latches once we've actually shown the overlay, so the brief frame where
   // gatewayState flips to "open" (connecting -> false) before the exit phase
   // kicks in doesn't unmount us and cause a flash.