feat: added performance statistics info when using embedded LLM

src/agent/core/AgentCore.ts

@@ -110,6 +110,7 @@ export class AgentCore {
         let assistantTextContent = '';
         const accumulatedToolCalls: ToolCall[] = [];
         let finishReason = 'stop';
+        let performanceInfo: StreamChunk['performanceInfo'];
 
         for await (const chunk of this.llmProvider.generateStream(conversationHistory, systemPrompt, tools)) {
           if (chunk.type === 'text') {
@@ -120,6 +121,7 @@ export class AgentCore {
             accumulatedToolCalls.push(chunk.toolCall);
           } else if (chunk.type === 'done') {
             finishReason = chunk.finishReason ?? 'stop';
+            performanceInfo = chunk.performanceInfo;
           }
         }
 
@@ -163,10 +165,9 @@ export class AgentCore {
           // LLM finished with text response (stop reason)
           this.agentState.updateMessage(this.currentAssistantMessageId, assistantTextContent);
           continueLoop = false;
+          yield { type: 'done', content: '', finishReason, performanceInfo };
         }
       }
-
-      yield { type: 'done', content: '', finishReason: 'stop' };
     } finally {
       this.currentUserMessageId = null;
       this.currentAssistantMessageId = null;

src/agent/llm/LocalBrowserLLMProvider.ts

@@ -296,6 +296,14 @@ export class LocalBrowserLLMProvider implements LLMProvider {
       yield { type: 'tool_call', content: '', toolCall };
     }
 
-    yield { type: 'done', content: '', finishReason };
+    yield {
+      type: 'done',
+      content: '',
+      finishReason,
+      performanceInfo: {
+        prefillTps,
+        generationTps,
+      },
+    };
   }
 }

src/agent/llm/StreamingTypes.ts

@@ -4,10 +4,16 @@
 
 import type { ToolCall } from '../core/AgentState';
 
+export interface PerformanceInfo {
+  prefillTps?: number;
+  generationTps?: number;
+}
+
 export interface StreamChunk {
   type: 'text' | 'tool_call' | 'tool_result' | 'done';
   content: string;
   toolCall?: ToolCall;
   toolResult?: { name: string; success: boolean; result: string };
-  finishReason?: string; // 'stop' | 'tool_calls' — present on 'done' chunks
+  performanceInfo?: PerformanceInfo;
+  finishReason?: string;
 }