Version or commit
16ab8e2 (origin/main); matrix pin open-telemetry/semantic-conventions-genai@a685613
Observed behavior
docs/otel-genai-compatibility.md says the pinned OTel convention defines a Histogram with input/output token types, then:
AgentTrust records additive, attribution-scoped facts and also distinguishes cache and reasoning categories.
That framing implies cache/reasoning are an AgentTrust-only distinction. At the matrix’s pinned upstream revision, OTel already defines span usage attributes including:
gen_ai.usage.cache_read.input_tokens
gen_ai.usage.cache_creation.input_tokens
gen_ai.usage.reasoning.output_tokens
while gen_ai.token.type for gen_ai.client.token.usage remains only input / output.
compatibility/otel-genai.json metrics note for agentrust.usage.tokens similarly says AgentTrust records “additional token categories,” which overstates the gap the same way.
Expected behavior
The docs should not implying that OpenTelemetry lacks cache/reasoning token fields.
Accurate distinction:
- OpenTelemetry’s token histogram type is only
input / output. Cache and reasoning tokens are tracked as separate span attributes in the pinned OTel version (they already exist upstream).
- What AgentTrust adds is attribution scope, labeling usage as belonging to a
model_call, agent_step, task, agent_run, or workflow_run, not “we invented cache/reasoning and OTel did not.”
Minimal reproduction
- Read https://github.com/agentrust-io/agentrust-telemetry/blob/16ab8e2/docs/otel-genai-compatibility.md#L31-L36
- Read the metrics note at https://github.com/agentrust-io/agentrust-telemetry/blob/16ab8e2/compatibility/otel-genai.json#L36
- At pin a685613a207a580163353b8e48a7ad88967e7b42, confirm OTel already names the cache/reasoning usage attributes, e.g. in gen-ai metrics/registry docs (
gen_ai.usage.cache_read.input_tokens, gen_ai.usage.cache_creation.input_tokens, gen_ai.usage.reasoning.output_tokens) while gen_ai.token.type remains input/output only.
Environment and additional context
Docs/compatibility-claim accuracy only, no schema or SDK wire change.
Version or commit
16ab8e2 (origin/main); matrix pin open-telemetry/semantic-conventions-genai@a685613
Observed behavior
docs/otel-genai-compatibility.mdsays the pinned OTel convention defines a Histogram withinput/outputtoken types, then:That framing implies cache/reasoning are an AgentTrust-only distinction. At the matrix’s pinned upstream revision, OTel already defines span usage attributes including:
gen_ai.usage.cache_read.input_tokensgen_ai.usage.cache_creation.input_tokensgen_ai.usage.reasoning.output_tokenswhile
gen_ai.token.typeforgen_ai.client.token.usageremains onlyinput/output.compatibility/otel-genai.jsonmetrics note foragentrust.usage.tokenssimilarly says AgentTrust records “additional token categories,” which overstates the gap the same way.Expected behavior
The docs should not implying that OpenTelemetry lacks cache/reasoning token fields.
Accurate distinction:
input/output. Cache and reasoning tokens are tracked as separate span attributes in the pinned OTel version (they already exist upstream).model_call,agent_step,task,agent_run, orworkflow_run, not “we invented cache/reasoning and OTel did not.”Minimal reproduction
gen_ai.usage.cache_read.input_tokens,gen_ai.usage.cache_creation.input_tokens,gen_ai.usage.reasoning.output_tokens) whilegen_ai.token.typeremainsinput/outputonly.Environment and additional context
Docs/compatibility-claim accuracy only, no schema or SDK wire change.