fix(core): use output-only tokens and accumulate across subagent rounds

Subagent token display had two bugs:
- Used totalTokenCount (input+output) instead of candidatesTokenCount
  (output-only), causing mixed units when aggregated with main stream
- Overwrote tokenCount per round instead of accumulating, so multi-round
  subagents only showed the last round's count

Co-Authored-By: Qwen-Coder <noreply@qwen.ai>

packages/core/src/tools/agent.ts

@@ -486,16 +486,21 @@ class AgentToolInvocation extends BaseToolInvocation<AgentParams, ToolResult> {
     });
 
     // Track real-time token consumption from subagent API calls.
-    // Each USAGE_METADATA event carries the cumulative totalTokenCount for the
+    // Each USAGE_METADATA event carries per-round usage, so we accumulate
-    // subagent session, so we replace (not accumulate) on every event.
+    // output tokens across rounds.  We use candidatesTokenCount (output-only)
+    // to stay consistent with the main stream's chars/4 output-token estimate.
+    let accumulatedOutputTokens = 0;
     this.eventEmitter.on(
       AgentEventType.USAGE_METADATA,
       (...args: unknown[]) => {
         const event = args[0] as AgentUsageEvent;
-        const total =
+        const outputTokens = event.usage?.candidatesTokenCount ?? 0;
-          event.usage?.totalTokenCount ?? event.usage?.promptTokenCount ?? 0;
+        if (outputTokens > 0) {
-        if (total > 0) {
+          accumulatedOutputTokens += outputTokens;
-          this.updateDisplay({ tokenCount: total }, updateOutput);
+          this.updateDisplay(
+            { tokenCount: accumulatedOutputTokens },
+            updateOutput,
+          );
         }
       },
     );