fix(chat): restrict thinking prefix to claude

Initialize the composer thinking mode from localStorage so reloads keep the user's

selected mode instead of falling back to Standard.

Keep the chosen mode after sending because the selector behaves like a

preference, not a one-shot modifier for a single prompt.

public/modelConstants.js

@@ -11,7 +11,7 @@ export const CLAUDE_MODELS = {
     {
       value: "default",
       label: "Default (recommended)",
-      description: "Use the default model (currently Opus 4.8 (1M context)) · $5/$25 per Mtok",
+      description: "Use the default model (currently Opus 4.7 (1M context)) · $5/$25 per Mtok",
     },
     {
       value: "sonnet",

server/claude-sdk.js

@@ -304,11 +304,7 @@ function extractTokenBudget(sdkMessage) {
 
   const messageUsage = sdkMessage.message?.usage || sdkMessage.usage;
   if (messageUsage && typeof messageUsage === 'object') {
-    const directInputTokens = readNumber(messageUsage.input_tokens ?? messageUsage.inputTokens);
+    const inputTokens = readNumber(messageUsage.input_tokens ?? messageUsage.inputTokens);
-    const cacheCreationTokens = readNumber(messageUsage.cache_creation_input_tokens ?? messageUsage.cacheCreationInputTokens ?? messageUsage.cacheCreationTokens);
-    const cacheReadTokens = readNumber(messageUsage.cache_read_input_tokens ?? messageUsage.cacheReadInputTokens ?? messageUsage.cacheReadTokens);
-    const cacheTokens = cacheCreationTokens + cacheReadTokens;
-    const inputTokens = directInputTokens + cacheTokens;
     const outputTokens = readNumber(messageUsage.output_tokens ?? messageUsage.outputTokens);
     const totalUsed = inputTokens + outputTokens;
     const contextWindow = parseInt(process.env.CONTEXT_WINDOW, 10) || 160000;
@@ -318,9 +314,6 @@ function extractTokenBudget(sdkMessage) {
       total: contextWindow,
       inputTokens,
       outputTokens,
-      cacheReadTokens,
-      cacheCreationTokens,
-      cacheTokens,
       breakdown: {
         input: inputTokens,
         output: outputTokens,

server/index.js

@@ -87,11 +87,6 @@ const installMode = fs.existsSync(path.join(APP_ROOT, '.git')) ? 'git' : 'npm';
 
 console.log('SERVER_PORT from env:', process.env.SERVER_PORT);
 
-function readUsageNumber(value) {
-    const parsed = Number(value);
-    return Number.isFinite(parsed) ? parsed : 0;
-}
-
 const app = express();
 const server = http.createServer(app);
 
@@ -1391,8 +1386,6 @@ app.get('/api/projects/:projectId/sessions/:sessionId/token-usage', authenticate
         const contextWindow = Number.isFinite(parsedContextWindow) ? parsedContextWindow : 160000;
         let inputTokens = 0;
         let outputTokens = 0;
-        let cacheReadTokens = 0;
-        let cacheCreationTokens = 0;
 
         // Find the latest assistant message with usage data (scan from end)
         for (let i = lines.length - 1; i >= 0; i--) {
@@ -1404,11 +1397,8 @@ app.get('/api/projects/:projectId/sessions/:sessionId/token-usage', authenticate
                     const usage = entry.message.usage;
 
                     // Use token counts from latest assistant message only
-                    const directInputTokens = readUsageNumber(usage.input_tokens ?? usage.inputTokens);
+                    inputTokens = usage.input_tokens || 0;
-                    cacheReadTokens = readUsageNumber(usage.cache_read_input_tokens ?? usage.cacheReadInputTokens ?? usage.cacheReadTokens);
+                    outputTokens = usage.output_tokens || 0;
-                    cacheCreationTokens = readUsageNumber(usage.cache_creation_input_tokens ?? usage.cacheCreationInputTokens ?? usage.cacheCreationTokens);
-                    inputTokens = directInputTokens + cacheReadTokens + cacheCreationTokens;
-                    outputTokens = readUsageNumber(usage.output_tokens ?? usage.outputTokens);
 
                     break; // Stop after finding the latest assistant message
                 }
@@ -1419,16 +1409,12 @@ app.get('/api/projects/:projectId/sessions/:sessionId/token-usage', authenticate
         }
 
         const totalUsed = inputTokens + outputTokens;
-        const cacheTokens = cacheReadTokens + cacheCreationTokens;
 
         res.json({
             used: totalUsed,
             total: contextWindow,
             inputTokens,
             outputTokens,
-            cacheReadTokens,
-            cacheCreationTokens,
-            cacheTokens,
             breakdown: {
                 input: inputTokens,
                 output: outputTokens

server/modules/websocket/services/websocket-server.service.ts

@@ -31,24 +31,6 @@ export function createWebSocketServer(
   });
 
   wss.on('connection', (ws, request) => {
-    // Keep WebSocket alive across reverse-proxy idle timeouts (Cloudflare ~100s,
-    // AWS ALB 60s, nginx 60s, etc.). Without app-level pings these connections
-    // are silently torn down even when the UI is active, causing repeated
-    // reconnect cycles. ws library heartbeat is opt-in.
-    const HEARTBEAT_INTERVAL_MS = 30_000;
-    const heartbeat = setInterval(() => {
-      if (ws.readyState === ws.OPEN) {
-        try {
-          ws.ping();
-        } catch {
-          // socket may have been closed concurrently — interval will be cleared below
-        }
-      }
-    }, HEARTBEAT_INTERVAL_MS);
-    const stopHeartbeat = () => clearInterval(heartbeat);
-    ws.on('close', stopHeartbeat);
-    ws.on('error', stopHeartbeat);
-
     const incomingRequest = request as AuthenticatedWebSocketRequest;
     const url = incomingRequest.url ?? '/';
     const pathname = new URL(url, 'http://localhost').pathname;

server/routes/agent.js

@@ -592,14 +592,12 @@ class ResponseCollector {
       }
     }
 
-    const inputTokens = totalInput + totalCacheRead + totalCacheCreation;
-
     return {
-      inputTokens,
+      inputTokens: totalInput,
       outputTokens: totalOutput,
       cacheReadTokens: totalCacheRead,
       cacheCreationTokens: totalCacheCreation,
-      totalTokens: inputTokens + totalOutput
+      totalTokens: totalInput + totalOutput + totalCacheRead + totalCacheCreation
     };
   }
 }

server/routes/commands.js

@@ -268,35 +268,16 @@ Custom commands can be created in:
           tokenUsage.contextWindow ??
           0,
       ) || 0;
-    const normalizedInputValue =
+    const inputTokensRaw =
-      tokenUsage.inputTokens ??
-      tokenUsage.input ??
-      tokenUsage.cumulativeInputTokens ??
-      tokenUsage.breakdown?.input ??
-      tokenUsage.promptTokens;
-    const directInputTokens =
       Number(
-        normalizedInputValue ??
+        tokenUsage.inputTokens ??
+          tokenUsage.input ??
           tokenUsage.input_tokens ??
-          0
+          tokenUsage.cumulativeInputTokens ??
-      ) || 0;
+          tokenUsage.breakdown?.input ??
-    const cacheReadTokens =
+          tokenUsage.promptTokens ??
-      Number(
-        tokenUsage.cacheReadTokens ??
-          tokenUsage.cache_read_input_tokens ??
-          tokenUsage.cacheReadInputTokens ??
           0,
       ) || 0;
-    const cacheCreationTokens =
-      Number(
-        tokenUsage.cacheCreationTokens ??
-          tokenUsage.cache_creation_input_tokens ??
-          tokenUsage.cacheCreationInputTokens ??
-          0,
-      ) || 0;
-    const inputTokens = normalizedInputValue == null
-      ? directInputTokens + cacheReadTokens + cacheCreationTokens
-      : directInputTokens;
     const outputTokens =
       Number(
         tokenUsage.outputTokens ??
@@ -307,9 +288,8 @@ Custom commands can be created in:
           tokenUsage.completionTokens ??
           0,
       ) || 0;
-    const computedUsed = inputTokens + outputTokens;
+    const hasTokenBreakdown = inputTokensRaw > 0 || outputTokens > 0;
-    const hasTokenBreakdown = computedUsed > 0;
+    const used = reportedUsed || inputTokensRaw + outputTokens;
-    const used = Math.max(reportedUsed, computedUsed);
 
     return {
       type: "builtin",
@@ -322,7 +302,7 @@ Custom commands can be created in:
         ...(hasTokenBreakdown
           ? {
               tokenBreakdown: {
-                input: inputTokens,
+                input: inputTokensRaw,
                 output: outputTokens,
               },
             }

src/components/chat/hooks/useChatComposerState.ts

@@ -143,6 +143,21 @@ const createFakeSubmitEvent = () => {
   return { preventDefault: () => undefined } as unknown as FormEvent<HTMLFormElement>;
 };
 
+const THINKING_MODE_STORAGE_KEY = 'chat-thinking-mode';
+
+const getInitialThinkingMode = () => {
+  if (typeof window === 'undefined') {
+    return 'none';
+  }
+
+  const savedMode = safeLocalStorage.getItem(THINKING_MODE_STORAGE_KEY);
+  if (!savedMode) {
+    return 'none';
+  }
+
+  return thinkingModes.some((mode) => mode.id === savedMode) ? savedMode : 'none';
+};
+
 const getNotificationSessionSummary = (
   selectedSession: ProjectSession | null,
   fallbackInput: string,
@@ -204,7 +219,7 @@ export function useChatComposerState({
   const [uploadingImages, setUploadingImages] = useState<Map<string, number>>(new Map());
   const [imageErrors, setImageErrors] = useState<Map<string, string>>(new Map());
   const [isTextareaExpanded, setIsTextareaExpanded] = useState(false);
-  const [thinkingMode, setThinkingMode] = useState('none');
+  const [thinkingMode, setThinkingMode] = useState(getInitialThinkingMode);
   const [commandModalPayload, setCommandModalPayload] = useState<CommandModalPayload | null>(null);
 
   const textareaRef = useRef<HTMLTextAreaElement>(null);
@@ -564,7 +579,7 @@ export function useChatComposerState({
 
       let messageContent = currentInput;
       const selectedThinkingMode = thinkingModes.find((mode: { id: string; prefix?: string }) => mode.id === thinkingMode);
-      if (selectedThinkingMode && selectedThinkingMode.prefix) {
+      if (provider === 'claude' && selectedThinkingMode && selectedThinkingMode.prefix) {
         messageContent = `${selectedThinkingMode.prefix}: ${currentInput}`;
       }
 
@@ -749,7 +764,6 @@ export function useChatComposerState({
       setUploadingImages(new Map());
       setImageErrors(new Map());
       setIsTextareaExpanded(false);
-      setThinkingMode('none');
 
       if (textareaRef.current) {
         textareaRef.current.style.height = 'auto';
@@ -795,6 +809,10 @@ export function useChatComposerState({
     inputValueRef.current = input;
   }, [input]);
 
+  useEffect(() => {
+    safeLocalStorage.setItem(THINKING_MODE_STORAGE_KEY, thinkingMode);
+  }, [thinkingMode]);
+
   useEffect(() => {
     if (!selectedProjectId) {
       return;

src/components/chat/view/subcomponents/ChatComposer.tsx

@@ -295,7 +295,6 @@ export default function ChatComposer({
 
             <PromptInputTextarea
               ref={textareaRef}
-              dir="auto"
               value={input}
               onChange={onInputChange}
               onClick={onTextareaClick}

src/components/chat/view/subcomponents/MessageComponent.tsx

@@ -120,7 +120,7 @@ const MessageComponent = memo(({ message, prevMessage, createDiff, onFileOpen, o
         /* User message bubble on the right */
         <div className="flex w-full items-end space-x-0 sm:w-auto sm:max-w-[85%] sm:space-x-3 md:max-w-md lg:max-w-lg xl:max-w-xl">
           <div className="group flex-1 rounded-2xl rounded-br-md bg-blue-600 px-3 py-2 text-white shadow-sm sm:flex-initial sm:px-4">
-            <div dir="auto" className="whitespace-pre-wrap break-words text-sm">
+            <div className="whitespace-pre-wrap break-words text-sm">
               {message.content}
             </div>
             {message.images && message.images.length > 0 && (
@@ -405,7 +405,7 @@ const MessageComponent = memo(({ message, prevMessage, createDiff, onFileOpen, o
                 </ReasoningContent>
               </Reasoning>
             ) : (
-              <div dir="auto" className="text-sm text-gray-700 dark:text-gray-300">
+              <div className="text-sm text-gray-700 dark:text-gray-300">
                 {/* Reasoning accordion */}
                 {showThinking && message.reasoning && (
                   <Reasoning className="mb-3" defaultOpen={false}>

src/contexts/WebSocketContext.tsx

@@ -36,12 +36,8 @@ const useWebSocketProviderState = (): WebSocketContextType => {
   const { token } = useAuth();
 
   useEffect(() => {
-    // The cleanup below sets unmountedRef = true. Without this reset, every
-    // re-run of the effect (e.g. on token refresh) would short-circuit connect()
-    // at its unmounted guard and leave the socket permanently disconnected.
-    unmountedRef.current = false;
     connect();
-
+    
     return () => {
       unmountedRef.current = true;
       if (reconnectTimeoutRef.current) {

vite.config.js

@@ -37,10 +37,6 @@ export default defineConfig(({ mode }) => {
         '/shell': {
           target: `ws://${proxyHost}:${serverPort}`,
           ws: true
-        },
-        '/plugin-ws': {
-          target: `ws://${proxyHost}:${serverPort}`,
-          ws: true
         }
       }
     },