Attribute conventions

The observability ecosystem hasn't agreed on attribute names for LLM telemetry yet. Langfuse and SigNoz use the (pre-standard) gen_ai.* convention from OTel's GenAI SIG; Phoenix uses Arize's OpenInference with llm.token_count.* and input.value / output.value.

hermes-otel emits both conventions on the same span so whichever backend reads from it picks up the data it's expecting. No vendor-specific adapter code per backend.

Token counts (on api.* spans)

Metric	Langfuse / gen_ai	Phoenix / OpenInference
Prompt tokens	gen_ai.usage.input_tokens	llm.token_count.prompt
Completion tokens	gen_ai.usage.output_tokens	llm.token_count.completion
Total tokens	gen_ai.usage.total_tokens	llm.token_count.total
Cache read	gen_ai.usage.cache_read_input_tokens	llm.token_count.prompt_details.cache_read
Cache write	gen_ai.usage.cache_creation_input_tokens	llm.token_count.prompt_details.cache_write
Reasoning	gen_ai.usage.reasoning.output_tokens	llm.token_count.completion_details.reasoning

Cache read/write are only populated when the provider reports them (Anthropic's prompt caching, OpenAI's — both surface them in their API responses). Reasoning tokens are emitted only by reasoning-capable models and are a subset of completion/output tokens (already counted in the completion and total figures), surfaced separately for visibility — don't add them on top of the total.

Message content (on llm.* spans)

	Langfuse / gen_ai	Phoenix / OpenInference
User message	gen_ai.input.messages (message JSON)	input.value
Assistant response	gen_ai.output.messages (message JSON)	output.value
Content type	(not set)	input.mime_type, output.mime_type

When conversation capture is on, input.value and gen_ai.input.messages become JSON of the full message list, input.mime_type becomes application/json, and hermes.conversation.message_count records how many messages were passed.

Model / request metadata

	Langfuse / gen_ai	Phoenix / OpenInference
Model name	gen_ai.request.model	llm.model_name
Provider	gen_ai.provider.name and the older gen_ai.system	llm.provider
Request parameters	gen_ai.request.temperature, gen_ai.request.max_tokens, gen_ai.request.top_p, … (on api.*)	llm.request.max_tokens
Finish reason	gen_ai.response.finish_reasons (an array)	llm.response.finish_reason

Request parameters are set only when Hermes passes them to the hook.

Tool spans

Tool spans carry both conventions too:

Attribute	Meaning
tool.name / gen_ai.tool.name	Tool name
gen_ai.tool.call.id	The model's tool-call id (also on the approval.* span that gated it)
input.value / gen_ai.tool.call.arguments	Tool args (JSON string)
output.value / gen_ai.tool.call.result	Tool result (string)
hermes.tool.target	Inferred file / URL (plugin-specific)
hermes.tool.command	Inferred shell command (plugin-specific)
hermes.tool.outcome	completed / error / timeout / blocked / cancelled (plugin-specific)
hermes.skill.name	Inferred skill name (plugin-specific, optional)

See Tool identity.

Session / turn metadata (on the agent / cron root)

Attribute	Meaning
hermes.session.kind	cli / telegram / discord / cron / ...
hermes.session.id	Hermes session ID
session.id	Standard OTel alias of the above (gen_ai.conversation.id is the GenAI one)
user.id	platform:sender on gateway sessions, only with capture_sender_id: true
hermes.turn.*	Turn summary (see Turn summary)

Resource-level attributes

Set on the OTel Resource and therefore stamped on every span:

Attribute	Source
service.name	hermes-agent, or the profile name with service_name_from_profile (override via resource_attributes:; OTEL_SERVICE_NAME has no effect)
service.instance.id	A UUID per process, so two Hermes processes never share a metric series
service.version	hermes-otel plugin version
process.pid	Process id
hermes.profile	The Hermes profile this process serves (also on the root span)
hermes.kanban.task_id, hermes.kanban.board, hermes.kanban.run_id	A dispatcher-spawned kanban worker's card, board and attempt; absent elsewhere
openinference.project.name	project_name / OTEL_PROJECT_NAME (the Phoenix project)
host.name	Only with host_metrics: true
plus any resource_attributes: (and the deprecated global_tags:) from hermes_otel.yaml

The full list with every source is in the span attributes reference.

common_attributes: is the per-signal counterpart: its keys are set on every span, every log record and every metric point as plain attributes (never on the Resource), for backends that do not index resource attributes per signal. See Which attributes reach which signal.

Why dual-convention rather than pick one?

Every backend supports a different set. Emitting both is cheap (same span, a few extra key-value pairs) and saves every user from writing their own mapping adapter. When the GenAI spec stabilises and Phoenix/Langfuse converge, this will simplify.

Roadmap

• OpenInference Tool span kind is now stable — already emitted.
• The gen_ai.tool.* names are emitted already; they will follow the GenAI spec as it changes.
• session.id is the standard OTel name; the plugin emits both hermes.session.id (for compatibility with older backends that key on it) and session.id. The former may be dropped in a future major version — watch the changelog.