align compression token estimates (#749)

tests/server/run-chat-usage.test.ts

@@ -0,0 +1,33 @@
+import { describe, expect, it } from 'vitest'
+import { countTokens } from '../../packages/server/src/lib/context-compressor'
+import { estimateUsageTokensFromMessages } from '../../packages/server/src/services/hermes/run-chat/usage'
+
+describe('run-chat usage token estimates', () => {
+  it('counts message content instead of serialized message payloads', () => {
+    const messages = [
+      { role: 'user', content: 'hello from user' },
+      { role: 'assistant', content: 'hello from assistant' },
+    ]
+
+    const usage = estimateUsageTokensFromMessages(messages)
+
+    expect(usage.inputTokens).toBe(countTokens('hello from user'))
+    expect(usage.outputTokens).toBe(countTokens('hello from assistant'))
+    expect(usage.inputTokens + usage.outputTokens).toBeLessThan(countTokens(JSON.stringify(messages)))
+  })
+
+  it('keeps assistant tool call tokens on the output side', () => {
+    const messages = [
+      {
+        role: 'assistant',
+        content: 'calling tool',
+        tool_calls: [{ id: 'call_1', type: 'function', function: { name: 'lookup', arguments: '{"q":"x"}' } }],
+      },
+    ]
+
+    const usage = estimateUsageTokensFromMessages(messages)
+
+    expect(usage.inputTokens).toBe(0)
+    expect(usage.outputTokens).toBe(countTokens('calling tool') + countTokens(String(messages[0].tool_calls || '')))
+  })
+})