diff --git a/frontend/src/components/StatusBar.jsx b/frontend/src/components/StatusBar.jsx
index 8007de2..9765531 100644
--- a/frontend/src/components/StatusBar.jsx
+++ b/frontend/src/components/StatusBar.jsx
@@ -23,6 +23,7 @@ export function StatusBar({ sessionStats, isProcessing, connected, permissionMod
     cacheCreationTokens = 0,
     numTurns = 0,
     contextWindow = 200000,
+    contextLeftPercent = null, // From Claude's "Context left until auto-compact: X%" message
     isCompacting = false,
   } = sessionStats || {};
 
@@ -30,12 +31,9 @@ export function StatusBar({ sessionStats, isProcessing, connected, permissionMod
   const currentMode = PERMISSION_MODES.find(m => m.value === permissionMode) || PERMISSION_MODES[0];
   const ModeIcon = currentMode.icon;
 
-  // Calculate context usage
-  // inputTokens from modelUsage represents tokens in current context (including cache reads)
-  // contextWindow is the model's max context size (e.g., 200000 for opus)
-  const contextUsed = inputTokens + outputTokens;
-  const contextPercent = contextWindow > 0 ? Math.min(100, (contextUsed / contextWindow) * 100) : 0;
-  const contextRemaining = Math.max(0, 100 - contextPercent);
+  // Context remaining: Use Claude's reported value if available, otherwise don't show
+  // Claude only reports this when context is getting low (< ~15%)
+  const contextRemaining = contextLeftPercent;
 
   // Format cost
   const formatCost = (cost) => {
@@ -139,43 +137,19 @@ export function StatusBar({ sessionStats, isProcessing, connected, permissionMod
 
         {/* Right side: Token usage */}
         <div className="flex items-center gap-4">
-          {/* Token counts */}
-          {contextUsed > 0 && (
-            <div className="flex items-center gap-3 text-dark-400">
-              <span className="flex items-center gap-1">
-                <span className="text-dark-500">In:</span>
-                <span className="text-cyan-400">{formatTokens(inputTokens)}</span>
-              </span>
-              <span className="flex items-center gap-1">
-                <span className="text-dark-500">Out:</span>
-                <span className="text-green-400">{formatTokens(outputTokens)}</span>
-              </span>
-              {cacheReadTokens > 0 && (
-                <span className="flex items-center gap-1">
-                  <Database className="w-3 h-3 text-purple-400" />
-                  <span className="text-purple-400">{formatTokens(cacheReadTokens)}</span>
-                </span>
-              )}
-            </div>
-          )}
-
-          {/* Context status - shows remaining context percentage */}
-          <div className="flex items-center gap-2">
-            <span className="text-dark-500">Context:</span>
-            {contextUsed > 0 ? (
+          {/* Context status - only shown when Claude reports it (context getting low) */}
+          {contextRemaining !== null && (
+            <div className="flex items-center gap-2">
+              <span className="text-dark-500">Context:</span>
               <span className={`${
                 contextRemaining <= 5 ? 'text-red-400 font-medium animate-pulse' :
                 contextRemaining <= 15 ? 'text-red-400' :
                 contextRemaining <= 30 ? 'text-yellow-400' : 'text-green-400'
               }`}>
-                {contextRemaining <= 30
-                  ? `${contextRemaining.toFixed(0)}% left`
-                  : `${contextPercent.toFixed(0)}% used`}
+                {contextRemaining}% left
               </span>
-            ) : (
-              <span className="text-green-400">ok</span>
-            )}
-          </div>
+            </div>
+          )}
         </div>
       </div>
     </div>
diff --git a/frontend/src/hooks/useClaudeSession.js b/frontend/src/hooks/useClaudeSession.js
index 59095d6..d2eee62 100644
--- a/frontend/src/hooks/useClaudeSession.js
+++ b/frontend/src/hooks/useClaudeSession.js
@@ -33,6 +33,7 @@ export function useClaudeSession() {
     cacheCreationTokens: 0,
     numTurns: 0,
     contextWindow: 200000, // Default, updated from modelUsage
+    contextLeftPercent: null, // From Claude's "Context left until auto-compact: X%" message
     isCompacting: false,
   });
 
@@ -371,18 +372,38 @@ export function useClaudeSession() {
           || Object.keys(modelUsage)[0];
         const primaryUsage = primaryModel ? modelUsage[primaryModel] : null;
 
-        // Calculate effective input tokens (what's actually in context)
-        // This is: new input + cache read (cache creation doesn't count towards context limit)
-        const effectiveInput = (event.usage?.input_tokens || 0) +
-                               (event.usage?.cache_read_input_tokens || 0);
+        // Context calculation: Use top-level usage (aggregated) as primary source
+        // Per GitHub issue anthropics/claude-agent-sdk-typescript#66:
+        // Cache tokens DON'T count toward the 200K context window - they're handled separately
+        // Only input_tokens + output_tokens count toward the context window
+        // Cache tokens can appear in the millions because they're cumulative across the session
+        const usage = event.usage || {};
+
+        // Prefer top-level usage (aggregated), fallback to primaryUsage (per-model)
+        const inputTokens = usage.input_tokens ?? primaryUsage?.inputTokens ?? 0;
+        const cacheRead = usage.cache_read_input_tokens ?? primaryUsage?.cacheReadInputTokens ?? 0;
+        const cacheCreation = usage.cache_creation_input_tokens ?? primaryUsage?.cacheCreationInputTokens ?? 0;
+        const outputTokens = usage.output_tokens ?? primaryUsage?.outputTokens ?? 0;
+
+        // Context = input + output tokens ONLY (cache tokens are separate)
+        const totalContextUsed = inputTokens + outputTokens;
+
+        console.log('Result event stats:', {
+          inputTokens,
+          cacheRead,
+          cacheCreation,
+          totalContextUsed,
+          outputTokens,
+          contextWindow: primaryUsage?.contextWindow
+        });
 
         setSessionStats(prev => ({
           totalCost: event.total_cost_usd ?? prev.totalCost,
-          // Use modelUsage for accurate per-turn counts, fallback to usage
-          inputTokens: primaryUsage?.inputTokens ?? effectiveInput ?? prev.inputTokens,
-          outputTokens: primaryUsage?.outputTokens ?? event.usage?.output_tokens ?? prev.outputTokens,
-          cacheReadTokens: primaryUsage?.cacheReadInputTokens ?? event.usage?.cache_read_input_tokens ?? prev.cacheReadTokens,
-          cacheCreationTokens: primaryUsage?.cacheCreationInputTokens ?? event.usage?.cache_creation_input_tokens ?? prev.cacheCreationTokens,
+          // Context = input + output tokens only (cache tokens don't count toward window)
+          inputTokens: totalContextUsed,
+          outputTokens: outputTokens,
+          cacheReadTokens: cacheRead,
+          cacheCreationTokens: cacheCreation,
           numTurns: event.num_turns ?? prev.numTurns,
           contextWindow: primaryUsage?.contextWindow ?? prev.contextWindow,
           isCompacting: false,
@@ -390,9 +411,22 @@ export function useClaudeSession() {
       }
     } else if (event.type === 'system' && event.subtype === 'result') {
       setIsProcessing(false);
-    } else if (event.type === 'system' && event.message?.includes?.('compact')) {
+    } else if (event.type === 'system' && event.message) {
+      // Parse "Context left until auto-compact: X%" message
+      const contextMatch = event.message.match(/Context left until auto-compact:\s*(\d+)%/i);
+      if (contextMatch) {
+        const contextLeft = parseInt(contextMatch[1], 10);
+        console.log('Context left from Claude:', contextLeft + '%');
+        setSessionStats(prev => ({
+          ...prev,
+          contextLeftPercent: contextLeft,
+          isCompacting: false,
+        }));
+      }
       // Detect compacting
-      setSessionStats(prev => ({ ...prev, isCompacting: true }));
+      if (event.message.includes('compact')) {
+        setSessionStats(prev => ({ ...prev, isCompacting: true }));
+      }
     }
   }, []);