Contact Details
k.kalidas@gmail.com
π¦ Package Version
0.4.21
ποΈ Framework Version
google adk
π Describe the Bug
AgentOps captures ADK Gemini LLM spans and latency correctly, but trace metrics show zero tokens and zero cost. The issue appears to be a token attribute naming mismatch. Google ADK emits token usage as gen_ai.usage.input_tokens and gen_ai.usage.output_tokens, while AgentOps appears to aggregate only gen_ai.usage.prompt_tokens, gen_ai.usage.completion_tokens, and gen_ai.usage.total_tokens. The ADK trace contains valid Gemini usage_metadata and non-zero token attributes, but AgentOps metrics ignore them.
Details:
Agent Framework: google-adk (Version: 2.1.0)
LLM : gemini-2.5-flash-lite
agentops (0.4.21)
Programming Language : Python 3.13
In AgentOps SpanAttributes, token usage constants are defined as:
LLM_USAGE_COMPLETION_TOKENS = "gen_ai.usage.completion_tokens"
LLM_USAGE_PROMPT_TOKENS = "gen_ai.usage.prompt_tokens"
LLM_USAGE_TOTAL_TOKENS = "gen_ai.usage.total_tokens"
But ADK emits:
gen_ai.usage.input_tokens
gen_ai.usage.output_tokens
This creates a mismatch:
ADK emits β gen_ai.usage.input_tokens
AgentOps expects β gen_ai.usage.prompt_tokens
ADK emits β gen_ai.usage.output_tokens
AgentOps expects β gen_ai.usage.completion_tokens
This appears to be why AgentOps trace metrics return zero tokens/cost
Please normalize input_tokens β prompt_tokens, output_tokens β completion_tokens, and compute total_tokens when absent.
adk single invocation.json
π€ Contribution
Contact Details
k.kalidas@gmail.com
π¦ Package Version
0.4.21
ποΈ Framework Version
google adk
π Describe the Bug
AgentOps captures ADK Gemini LLM spans and latency correctly, but trace metrics show zero tokens and zero cost. The issue appears to be a token attribute naming mismatch. Google ADK emits token usage as gen_ai.usage.input_tokens and gen_ai.usage.output_tokens, while AgentOps appears to aggregate only gen_ai.usage.prompt_tokens, gen_ai.usage.completion_tokens, and gen_ai.usage.total_tokens. The ADK trace contains valid Gemini usage_metadata and non-zero token attributes, but AgentOps metrics ignore them.
Details:
Agent Framework: google-adk (Version: 2.1.0)
LLM : gemini-2.5-flash-lite
agentops (0.4.21)
Programming Language : Python 3.13
In AgentOps SpanAttributes, token usage constants are defined as:
LLM_USAGE_COMPLETION_TOKENS = "gen_ai.usage.completion_tokens"
LLM_USAGE_PROMPT_TOKENS = "gen_ai.usage.prompt_tokens"
LLM_USAGE_TOTAL_TOKENS = "gen_ai.usage.total_tokens"
But ADK emits:
gen_ai.usage.input_tokens
gen_ai.usage.output_tokens
This creates a mismatch:
ADK emits β gen_ai.usage.input_tokens
AgentOps expects β gen_ai.usage.prompt_tokens
ADK emits β gen_ai.usage.output_tokens
AgentOps expects β gen_ai.usage.completion_tokens
This appears to be why AgentOps trace metrics return zero tokens/cost
Please normalize input_tokens β prompt_tokens, output_tokens β completion_tokens, and compute total_tokens when absent.
adk single invocation.json
π€ Contribution