fix: include Claude cache tokens in usage

Add dir="auto" to chat message content and composer textarea so
Persian and Arabic text automatically renders right-to-left
while English and other LTR text remains unaffected.

Co-authored-by: Haile <118998054+blackmammoth@users.noreply.github.com>

public/modelConstants.js

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

server/claude-sdk.js

@@ -304,7 +304,11 @@ function extractTokenBudget(sdkMessage) {
 
   const messageUsage = sdkMessage.message?.usage || sdkMessage.usage;
   if (messageUsage && typeof messageUsage === 'object') {
-    const inputTokens = readNumber(messageUsage.input_tokens ?? messageUsage.inputTokens);
+    const directInputTokens = 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;
@@ -314,6 +318,9 @@ function extractTokenBudget(sdkMessage) {
       total: contextWindow,
       inputTokens,
       outputTokens,
+      cacheReadTokens,
+      cacheCreationTokens,
+      cacheTokens,
       breakdown: {
         input: inputTokens,
         output: outputTokens,

server/index.js

@@ -87,6 +87,11 @@ 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);
 
@@ -1386,6 +1391,8 @@ 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--) {
@@ -1397,8 +1404,11 @@ app.get('/api/projects/:projectId/sessions/:sessionId/token-usage', authenticate
                     const usage = entry.message.usage;
 
                     // Use token counts from latest assistant message only
-                    inputTokens = usage.input_tokens || 0;
-                    outputTokens = usage.output_tokens || 0;
+                    const directInputTokens = readUsageNumber(usage.input_tokens ?? usage.inputTokens);
+                    cacheReadTokens = readUsageNumber(usage.cache_read_input_tokens ?? usage.cacheReadInputTokens ?? usage.cacheReadTokens);
+                    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
                 }
@@ -1409,12 +1419,16 @@ 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-auth.service.ts

@@ -20,13 +20,7 @@ export function verifyWebSocketClient(
   dependencies: WebSocketAuthDependencies
 ): boolean {
   const request = info.req as AuthenticatedWebSocketRequest;
-  const upgradeUrl = new URL(request.url ?? '/', 'http://localhost');
-  const loggedUrl = new URL(upgradeUrl);
-  if (loggedUrl.searchParams.has('token')) {
-    loggedUrl.searchParams.set('token', 'REDACTED');
-  }
-
-  console.log('WebSocket connection attempt to:', `${loggedUrl.pathname}${loggedUrl.search}`);
+  console.log('WebSocket connection attempt to:', request.url);
 
   // Platform mode: use the first DB user and skip token checks.
   if (dependencies.isPlatform) {
@@ -42,6 +36,7 @@ export function verifyWebSocketClient(
   }
 
   // OSS mode: read JWT from query string first, then Authorization header.
+  const upgradeUrl = new URL(request.url ?? '/', 'http://localhost');
   const token =
     upgradeUrl.searchParams.get('token') ??
     request.headers.authorization?.split(' ')[1] ??

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

@@ -31,6 +31,24 @@ 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

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

server/routes/commands.js

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

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

@@ -295,6 +295,7 @@ 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 className="whitespace-pre-wrap break-words text-sm">
+            <div dir="auto" 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 className="text-sm text-gray-700 dark:text-gray-300">
+              <div dir="auto" 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

@@ -36,6 +36,10 @@ 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 () => {

vite.config.js

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