fix(context): use normalized prompt size for usage meter

Upgrade the Qoder Agent SDK to 1.0.51 so the runtime can provide authoritative context usage. Keep the stream fallback accurate by treating input_tokens as the complete prompt size rather than adding cache detail fields that are already subsets.

Author: luoxuanzao

CHANGELOG.md

@@ -11,6 +11,12 @@ version with its date and start a fresh empty `[Unreleased]` above it.
 
 ## [Unreleased]
 
+### Fixed
+
+- The context usage meter no longer counts cached prompt tokens twice, so its
+  percentage now reflects the actual occupied portion of the selected context
+  window.
+
 ## [1.0.17] - 2026-10-10
 
 ### Changed

package-lock.json

@@ -56,7 +56,7 @@
     "@codemirror/state": "^6.5.0",
     "@codemirror/view": "^6.38.6",
     "@modelcontextprotocol/sdk": "~1.32.1",
-    "@qoder-ai/qoder-agent-sdk": "^1.0.45",
+    "@qoder-ai/qoder-agent-sdk": "^1.0.51",
     "tslib": "^2.8.1"
   },
   "overrides": {

package.json

@@ -165,8 +165,8 @@ export type StreamChunk =
 /**
  * Context window usage information.
  *
- * `contextTokens` is the Qoder-computed total token count in the context window:
- * `inputTokens + cacheCreationInputTokens + cacheReadInputTokens`.
+ * `contextTokens` is the complete prompt size reported by Qoder. Cache creation
+ * and cache read counts are subsets of `inputTokens`, not additional context.
  *
  * Cache token fields are optional for older Qoder CLI responses. Feature code
  * should use `contextTokens` for display, not recompute from the breakdown.

src/core/types/chat.ts

@@ -311,7 +311,9 @@ function toPromptUsageSnapshot(usage: MessageUsage): PromptUsageSnapshot {
     inputTokens,
     cacheCreationInputTokens,
     cacheReadInputTokens,
-    contextTokens: inputTokens + cacheCreationInputTokens + cacheReadInputTokens,
+    // qodercli normalizes input_tokens to the complete prompt size. Cache read
+    // and creation counts are subsets used for billing detail, not extra context.
+    contextTokens: inputTokens,
   };
 }
 
@@ -327,7 +329,7 @@ function mergePromptUsage(
     inputTokens,
     cacheCreationInputTokens,
     cacheReadInputTokens,
-    contextTokens: inputTokens + cacheCreationInputTokens + cacheReadInputTokens,
+    contextTokens: inputTokens,
   };
 }
 

src/qoder/stream/transform-qoder-message.ts

@@ -1522,17 +1522,17 @@ describe('transformSDKMessage', () => {
   });
 
   describe('assistant message usage extraction', () => {
-    it('yields usage info from main agent assistant message', () => {
+    it('does not double-count cache details toward context occupancy', () => {
       const message = msg({
         type: 'assistant',
         parent_tool_use_id: null, // Main agent
         message: {
           content: [{ type: 'text', text: 'Hello' }],
           usage: {
-            input_tokens: 1000,
+            input_tokens: 100_000,
             output_tokens: 500,
-            cache_creation_input_tokens: 300,
-            cache_read_input_tokens: 200,
+            cache_creation_input_tokens: 10_000,
+            cache_read_input_tokens: 90_000,
           },
         },
       });
@@ -1543,12 +1543,12 @@ describe('transformSDKMessage', () => {
       expect(usageResults).toHaveLength(1);
 
       const usage = (usageResults[0] as any).usage;
-      expect(usage.inputTokens).toBe(1000);
-      expect(usage.cacheCreationInputTokens).toBe(300);
-      expect(usage.cacheReadInputTokens).toBe(200);
-      expect(usage.contextTokens).toBe(1500); // 1000 + 300 + 200
-      expect(usage.contextWindow).toBe(200000); // Standard context window
-      expect(usage.percentage).toBe(1); // 1500 / 200000 * 100 rounded
+      expect(usage.inputTokens).toBe(100_000);
+      expect(usage.cacheCreationInputTokens).toBe(10_000);
+      expect(usage.cacheReadInputTokens).toBe(90_000);
+      expect(usage.contextTokens).toBe(100_000);
+      expect(usage.contextWindow).toBe(200_000);
+      expect(usage.percentage).toBe(50);
     });
 
     it('yields usage from assistant message when usage state has no stream prompt usage', () => {
@@ -1580,7 +1580,7 @@ describe('transformSDKMessage', () => {
         cacheCreationInputTokens: 300,
         cacheReadInputTokens: 200,
         contextWindow: 200000,
-        contextTokens: 1500,
+        contextTokens: 1000,
         percentage: 1,
       });
     });