fix(service-ai): estimate non-zero token usage in MemoryLLMAdapter

The in-memory adapter stands in for a real LLM in dev/CI/local-E2E, but
returned a flat `usage: {0,0,0}` on every path. That made every
token-metering feature — quota guardrails, lifetime caps, usage
dashboards, cost stops — impossible to exercise without a paid provider
key, i.e. without spending real money.

Estimate usage crudely (~4 chars/token, the standard ballpark) from the
prompt messages and echoed output across chat/complete/streamChat/
generateObject. Intentionally NOT provider-accurate — it exists so
usage-driven behaviour is testable money-free, not to bill anyone.

Adds a unit test asserting non-zero, additive (total = prompt+completion),
monotonic-in-input-size usage.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: os-zhuang

packages/services/service-ai/src/__tests__/ai-service.test.ts

@@ -45,6 +45,21 @@ describe('MemoryLLMAdapter', () => {
     expect(result.usage).toBeDefined();
   });
 
+  it('estimates non-zero, monotonic token usage so metering is testable money-free', async () => {
+    const short = await adapter.chat([{ role: 'user', content: 'hi' }]);
+    const long = await adapter.chat([
+      { role: 'system', content: 'You are a helpful assistant with a long preamble. '.repeat(20) },
+      { role: 'user', content: 'Tell me everything about my data. '.repeat(20) },
+    ]);
+    // Non-zero (the whole point — a flat 0 made quota/guardrail features untestable).
+    expect(short.usage!.totalTokens).toBeGreaterThan(0);
+    // total = prompt + completion, and grows with input size (monotonic).
+    expect(short.usage!.totalTokens).toBe(
+      short.usage!.promptTokens + short.usage!.completionTokens,
+    );
+    expect(long.usage!.promptTokens).toBeGreaterThan(short.usage!.promptTokens);
+  });
+
   it('should handle no user message in chat()', async () => {
     const messages: ModelMessage[] = [{ role: 'system', content: 'System only' }];
     const result = await adapter.chat(messages);

packages/services/service-ai/src/adapters/memory-adapter.ts

@@ -18,6 +18,27 @@ import type { LLMAdapter } from '@objectstack/spec/contracts';
  * Always echoes back the last user message prefixed with "[memory] ".
  * Useful for unit tests, CI pipelines, and local dev without an LLM key.
  */
+/**
+ * Rough token estimate for the in-memory adapter. The memory adapter stands in
+ * for a real LLM in dev/CI/local-E2E; returning a flat `0` usage made every
+ * token-metering feature (quota guardrails, usage dashboards, cost caps)
+ * impossible to exercise without a paid provider key. This is intentionally
+ * crude — ~4 chars/token, the standard ballpark — NOT provider-accurate. It
+ * exists so usage-driven behaviour can be tested money-free, not to bill anyone.
+ */
+function estimateTokens(text: string): number {
+  return text ? Math.max(1, Math.ceil(text.length / 4)) : 0;
+}
+
+function estimateUsage(messages: ModelMessage[], output = ''): AIResult['usage'] {
+  const promptText = messages
+    .map(m => (typeof m.content === 'string' ? m.content : JSON.stringify(m.content)))
+    .join('\n');
+  const promptTokens = estimateTokens(promptText);
+  const completionTokens = estimateTokens(output);
+  return { promptTokens, completionTokens, totalTokens: promptTokens + completionTokens };
+}
+
 export class MemoryLLMAdapter implements LLMAdapter {
   readonly name = 'memory';
 
@@ -66,7 +87,7 @@ export class MemoryLLMAdapter implements LLMAdapter {
           return {
             content: '',
             model: options?.model ?? 'memory',
-            usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+            usage: estimateUsage(messages),
             toolCalls: [
               {
                 type: 'tool-call',
@@ -89,7 +110,7 @@ export class MemoryLLMAdapter implements LLMAdapter {
       return {
         content: '',
         model: options?.model ?? 'memory',
-        usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+        usage: estimateUsage(messages),
         toolCalls: [
           {
             type: 'tool-call',
@@ -126,13 +147,13 @@ export class MemoryLLMAdapter implements LLMAdapter {
         return {
           content: `[memory] action ${payload.action ?? ''} failed: ${payload.error}`,
           model: options?.model ?? 'memory',
-          usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+          usage: estimateUsage(messages),
         };
       }
       return {
         content: `[memory] ${payload.message ?? 'Action executed.'} (${payload.action ?? 'action'})`,
         model: options?.model ?? 'memory',
-        usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+        usage: estimateUsage(messages),
       };
     }
 
@@ -163,15 +184,15 @@ export class MemoryLLMAdapter implements LLMAdapter {
         return {
           content: `[memory] query_data failed: ${payload.error}`,
           model: options?.model ?? 'memory',
-          usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+          usage: estimateUsage(messages),
         };
       }
       const records = payload.records ?? [];
       const count = payload.count ?? records.length;
       return {
         content: `[memory] Found ${count} record${count === 1 ? '' : 's'} for "${userText}".`,
         model: options?.model ?? 'memory',
-        usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+        usage: estimateUsage(messages),
       };
     }
 
@@ -182,15 +203,16 @@ export class MemoryLLMAdapter implements LLMAdapter {
     return {
       content,
       model: options?.model ?? 'memory',
-      usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+      usage: estimateUsage(messages),
     };
   }
 
   async complete(prompt: string, options?: AIRequestOptions): Promise<AIResult> {
+    const content = `[memory] ${prompt}`;
     return {
-      content: `[memory] ${prompt}`,
+      content,
       model: options?.model ?? 'memory',
-      usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+      usage: estimateUsage([{ role: 'user', content: prompt }], content),
     };
   }
 
@@ -208,7 +230,7 @@ export class MemoryLLMAdapter implements LLMAdapter {
     yield {
       type: 'finish',
       finishReason: 'stop' as const,
-      totalUsage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+      totalUsage: estimateUsage(messages, result.content),
       rawFinishReason: 'stop',
     } as unknown as TextStreamPart<ToolSet>;
   }
@@ -310,7 +332,7 @@ export class MemoryLLMAdapter implements LLMAdapter {
         return {
           object: result.data,
           model: options?.model ?? 'memory',
-          usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+          usage: estimateUsage(messages),
         };
       }
     }