feat: instrument llmService to record calls via llmStatsService

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

server/src/llm/__tests__/llmService.test.ts

@@ -161,6 +161,56 @@ describe('llmService', () => {
     expect(service.activeModel()).toBe('free/model:free');
   });
 
+  it('records successful call to stats service', async () => {
+    setupEnv();
+    globalThis.fetch = vi.fn().mockResolvedValue({
+      ok: true,
+      json: () => Promise.resolve({
+        choices: [{ message: { content: 'ok' } }],
+        usage: { prompt_tokens: 100, completion_tokens: 50 },
+      }),
+    });
+
+    const mockStatsService = {
+      record: vi.fn(),
+      getStats: vi.fn(),
+      isTrackingEnabled: vi.fn().mockReturnValue(true),
+      setTrackingEnabled: vi.fn(),
+    };
+
+    const service = createLlmService(undefined, mockStatsService);
+    await service.generate('backstory', { npcName: 'X', stats: 'S:1' });
+    await vi.advanceTimersByTimeAsync(100);
+
+    expect(mockStatsService.record).toHaveBeenCalledWith(
+      'backstory', 100, 50, 'free/model:free', 0, false, 0,
+    );
+  });
+
+  it('records failed call to stats service', async () => {
+    setupEnv();
+    globalThis.fetch = vi.fn().mockResolvedValue({
+      ok: false,
+      status: 400,
+      statusText: 'Bad Request',
+    });
+
+    const mockStatsService = {
+      record: vi.fn(),
+      getStats: vi.fn(),
+      isTrackingEnabled: vi.fn().mockReturnValue(true),
+      setTrackingEnabled: vi.fn(),
+    };
+
+    const service = createLlmService(undefined, mockStatsService);
+    await service.generate('backstory', { npcName: 'X', stats: 'S:1' });
+    await vi.advanceTimersByTimeAsync(100);
+
+    expect(mockStatsService.record).toHaveBeenCalledWith(
+      'backstory', 0, 0, 'free/model:free', 0, true, 0,
+    );
+  });
+
   it('destroy cleans up switch-back timer', async () => {
     setupEnv({ fallback: 'paid/model' });
     const resetAt = Date.now() + 60000;

server/src/llm/llmService.ts

@@ -6,6 +6,7 @@ import { renderTemplate } from './promptTemplate.js';
 import { templates } from './templates.js';
 import { createTokenTracker, type TokenTracker } from './tokenTracker.js';
 import type { LogService } from '../services/logService.js';
+import type { LlmStatsService } from './llmStatsService.js';
 
 export interface LlmService {
   generate(templateName: string, variables: Record<string, string>): Promise<string | null>;
@@ -18,7 +19,7 @@ export interface LlmService {
   tokenUsage(): TokenTracker;
 }
 
-export function createLlmService(logService?: LogService): LlmService {
+export function createLlmService(logService?: LogService, statsService?: LlmStatsService): LlmService {
   const config = getLlmConfig();
 
   if (!config.enabled || !config.model) {
@@ -87,15 +88,19 @@ export function createLlmService(logService?: LogService): LlmService {
       if (isRateLimited(result)) {
         logService?.log('warning', 'LLM', 'Rate limited, switching to fallback');
         switchToFallback(result.resetAt);
+        statsService?.record(templateName, 0, 0, currentModel, 0, true, 0);
         return null;
       }
 
       if (isSuccess(result)) {
         counters.record(currentModel);
+        const inputTokens = result.usage?.promptTokens ?? 0;
+        const outputTokens = result.usage?.completionTokens ?? 0;
         if (result.usage) {
-          tokenTracker.record(templateName, result.usage.promptTokens, result.usage.completionTokens);
+          tokenTracker.record(templateName, inputTokens, outputTokens);
           console.log(tokenTracker.lastSummary());
         }
+        statsService?.record(templateName, inputTokens, outputTokens, currentModel, result.retries, false, 0);
         const totalRequests = Object.values(counters.getStats()).reduce((sum, s) => sum + s.total, 0);
         if (totalRequests % 100 === 0) {
           console.log(`[LLM] ${counters.getSummary()}`);
@@ -104,6 +109,8 @@ export function createLlmService(logService?: LogService): LlmService {
         return result.content;
       }
 
+      // null result = failure
+      statsService?.record(templateName, 0, 0, currentModel, 0, true, 0);
       return result;
     },