Module semconv 

OTel GenAI / MCP semantic-convention names shared by every OTel emission surface (OtelHandler and OtelTracingBackend) so the two cannot drift. Other crates (e.g. lc-agents) may also reference the pub constants here instead of hardcoding attribute strings.

Stability (T10, v0.23 — 2026 alignment)

The GenAI semconv is still published at Development stability: in 2026 the working group moved the docs to a dedicated repository (open-telemetry/semantic-conventions-genai) and renames still happen (notably gen_ai.usage.cache_creation.input_tokens → GEN_AI_USAGE_CACHE_WRITE). Centralizing every name in this module is deliberate: when the registry moves a name, one constant changes and both export surfaces follow. The provider-extension and agent/tool attributes below follow the 2026 draft (gen-ai-agent-spans.md, gen-ai-events.md, mcp.md); the core chat attributes (provider.name, operation.name, request/response model, usage, finish_reasons) follow the March 2025 stable release.

MCP client spans live in lc-mcp, which cannot depend on this crate; their mcp.* constants are mirrored in lc_mcp::instrument with a cross-reference comment.

Constants

ERROR_TYPE

Standard error attribute (OTel general semconv).

EVENT_INFERENCE_DETAILS

gen_ai.client.inference.operation.details (2026 events model, Opt-In).

GEN_AI_AGENT_NAME

gen_ai.agent.name — agent identity for invoke_agent / plan spans.

GEN_AI_CHOICE_INDEX

gen_ai.choice.index.

GEN_AI_INPUT_MESSAGES

Structured input messages JSON on EVENT_INFERENCE_DETAILS.

GEN_AI_MESSAGE_CONTENT

gen_ai.message.content — payload key of the legacy per-message span events (see EVENT_INFERENCE_DETAILS doc for why those are retained).

GEN_AI_OPERATION_NAME

gen_ai.operation.name (chat / execute_tool / invoke_agent / plan / retrieve …).

GEN_AI_OUTPUT_MESSAGES

Structured output messages JSON on EVENT_INFERENCE_DETAILS.

GEN_AI_PROVIDER_NAME

gen_ai.provider.name (stabilized rename of experimental gen_ai.system).

GEN_AI_REQUEST_MAX_TOKENS

gen_ai.request.max_tokens.

GEN_AI_REQUEST_MODEL

gen_ai.request.model.

GEN_AI_REQUEST_TEMPERATURE

gen_ai.request.temperature.

GEN_AI_RESPONSE_FINISH_REASONS

gen_ai.response.finish_reasons (string array, plural).

GEN_AI_RESPONSE_MODEL

gen_ai.response.model.

GEN_AI_TOOL_CALL_ARGUMENTS

gen_ai.tool.call.arguments — Opt-In: recorded only when explicitly enabled.

GEN_AI_TOOL_CALL_ID

gen_ai.tool.call.id — the provider tool-call id (Recommended, 2026 draft).

GEN_AI_TOOL_CALL_RESULT

gen_ai.tool.call.result — Opt-In: recorded only when explicitly enabled.

GEN_AI_TOOL_DESCRIPTION

gen_ai.tool.description (Recommended, 2026 draft).

GEN_AI_TOOL_NAME

gen_ai.tool.name (tool spans; conditionally required on MCP tool calls).

GEN_AI_USAGE_CACHE_READ

gen_ai.usage.cache_read.input_tokens (Anthropic prompt-cache hits).

GEN_AI_USAGE_CACHE_WRITE

gen_ai.usage.cache_write.input_tokens — the 2026-draft rename of gen_ai.usage.cache_creation.input_tokens (still the published name in registry v1.40). Provider payloads use both spellings (cache_write_input_tokens and the Anthropic legacy cache_creation_input_tokens); the handler accepts either key.

GEN_AI_USAGE_INPUT_TOKENS

gen_ai.usage.input_tokens.

GEN_AI_USAGE_OUTPUT_TOKENS

gen_ai.usage.output_tokens.

GEN_AI_USAGE_REASONING

gen_ai.usage.reasoning.output_tokens (reasoning/thinking tokens).

RUN_ID_ATTR

Framework join key: the run id (also carried on evaluation reports).

TRACE_ID_ATTR

Framework join key: the trace id a run belongs to.