fix: Fix incorrect token count mapping in telemetry

Merge #2109 Fixes #2105 ## Problem When integrating Google ADK with Langfuse using the @observe decorator, the usage details displayed in Langfuse web UI were incorrect. The root cause was in the telemetry implementation where total_token_count was being mapped to gen_ai.usage.output_tokens instead of candidates_token_count. - Expected mapping: - candidates_token_count → completion_tokens (output tokens) - prompt_token_count → prompt_tokens (input tokens) - Previous incorrect mapping: - total_token_count → completion_tokens (wrong!) - prompt_token_count → prompt_tokens (correct) ## Solution Updated trace_call_llm function in telemetry.py to use candidates_token_count for output token tracking instead of total_token_count, ensuring proper token count reporting to observability tools like Langfuse. ## Testing plan - Updated test expectations in test_telemetry.py - Verified telemetry tests pass - Manual verification with Langfuse integration ## Screenshots **Before** <img width="1187" height="329" alt="Screenshot from 2025-07-22 20-20-33" src="https://github.com/user-attachments/assets/ad5fc957-64a2-4524-bd31-0cebb15a5270" /> **After** <img width="1187" height="329" alt="Screenshot from 2025-07-22 20-21-40" src="https://github.com/user-attachments/assets/3920df2a-be75-47e0-9bd0-f961bb72c838" /> _Notes_: From the screenshot, there's another problem: thoughts_token_count field is not mapped, but this should be another issue imo COPYBARA_INTEGRATE_REVIEW=#2109 from tl-nguyen:fix-telemetry-token-count-mapping 3d043f5 PiperOrigin-RevId: 786827802

Author: tl-nguyen

src/google/adk/telemetry.py

@@ -202,7 +202,7 @@ def trace_call_llm(
  )
  span.set_attribute(
  'gen_ai.usage.output_tokens',
- llm_response.usage_metadata.total_token_count,
+ llm_response.usage_metadata.candidates_token_count,
  )
 
 

tests/unittests/test_telemetry.py

@@ -155,15 +155,17 @@ async def test_trace_call_llm_usage_metadata(monkeypatch, mock_span_fixture):
  llm_response = LlmResponse(
  turn_complete=True,
  usage_metadata=types.GenerateContentResponseUsageMetadata(
- total_token_count=100, prompt_token_count=50
+ total_token_count=100,
+ prompt_token_count=50,
+ candidates_token_count=50,
  ),
  )
  trace_call_llm(invocation_context, 'test_event_id', llm_request, llm_response)
 
  expected_calls = [
  mock.call('gen_ai.system', 'gcp.vertex.agent'),
  mock.call('gen_ai.usage.input_tokens', 50),
- mock.call('gen_ai.usage.output_tokens', 100),
+ mock.call('gen_ai.usage.output_tokens', 50),
  ]
  assert mock_span_fixture.set_attribute.call_count == 9
  mock_span_fixture.set_attribute.assert_has_calls(