OpenTelemetry GenAI semantic conventions

The OpenTelemetry semantic conventions for generative AI define the span names, span kinds, gen_ai.* attributes, events and metrics for model inference, embeddings, tool execution, agents and the Model Context Protocol. They are maintained in the open-telemetry/semantic-conventions-genai repository and have Development status. With this skill the agent instruments GenAI calls so that any OpenTelemetry backend can read them.

Follow the workflow below step by step. Every rule here comes from a source in Sources, with the section it cites. When a rule and the pinned source disagree, the source wins; when the source has a newer revision than the pin, follow the refresh steps.

Inputs (fill in, or ask before starting)

• Role: instrumentation author (library or framework), application developer adding manual spans, or backend consumer reading the data.
• Revision: Draft posture: build. The pin is commit b31e9e8ea26ac1c086d3313d474e31d7c3f391ae (2026-09-30) of semantic-conventions-genai, schema URL https://opentelemetry.io/schemas/gen-ai-dev/1.42.0-dev, built on semantic conventions v1.44.0. The repository has no releases or tags, so pin the commit and the schema URL.
• Content capture: whether prompts, completions, system instructions and tool arguments may be recorded, and where (spans, events, or external storage).
• Sources: when refreshing this skill, re-read every URL in Sources, check the repository for a tag or a new schema_url in model/manifest.yaml, and update the pins.

Invariants

1. Everything here is Development. The conventions, spans, attributes, events and metrics carry Development status (GenAI README; manifest stability: development). Names can change between commits; record the schema URL you implemented.
2. Existing instrumentations do not switch conventions silently. Instrumentations that emit v1.36.0 or earlier keep emitting it by default, and add OTEL_SEMCONV_STABILITY_OPT_IN with the value gen_ai_latest_experimental to emit the latest experimental conventions instead of the old ones (semantic conventions v1.41.1 GenAI README).
3. gen_ai.operation.name is Required on every GenAI span, and gen_ai.provider.name on inference, embeddings, create_agent and client invoke_agent spans. Use a well-known value when one applies; otherwise a custom value is allowed (spans, agent spans, registry).
4. Span names are low-cardinality: {gen_ai.operation.name} {gen_ai.request.model} for inference and embeddings, execute_tool {gen_ai.tool.name}, invoke_agent {gen_ai.agent.name}, create_agent {gen_ai.agent.name} (spans, agent spans).
5. Span kind follows the call: CLIENT for remote model, embeddings and agent-service calls (INTERNAL allowed for in-process models); INTERNAL for execute_tool and in-process invoke_agent (spans, agent spans).
6. A span covers the whole logical operation, including automatic retries, until the response is fully received or the operation ends in an error or cancellation (spans, Spans).
7. Content is not captured by default. gen_ai.system_instructions, gen_ai.input.messages, gen_ai.output.messages, gen_ai.tool.definitions, gen_ai.tool.call.arguments and gen_ai.tool.call.result are Opt-In; offer an option to enable them (spans, Capturing instructions, inputs, and outputs).
8. Errors set error.type, and span status follows the Recording Errors document (spans).
9. Token counts report billable tokens when a system reports both used and billable tokens, and the per-operation histograms are never used for totals or cost (token metrics).

Workflow

1. Choose the spans. Map each GenAI call in the code to an inference, embeddings, retrieval, execute_tool, create_agent, invoke_agent, invoke_workflow or MCP span. -> references/spans.md ✓ Each call site has a span type, a name pattern and a span kind.
2. Set the attributes. Fill the Required and Conditionally Required attributes, then the Recommended ones the response provides, such as usage and finish reasons. -> references/attributes.md ✓ No span lacks gen_ai.operation.name or (where required) gen_ai.provider.name.
3. Decide on content capture. Default off; if enabled, choose span attributes or external storage, and apply truncation. -> references/attributes.md ✓ A configuration flag controls capture and is off by default.
4. Record metrics and events. Emit gen_ai.client.operation.duration with the advised buckets, the token usage counters, and the exception and evaluation events where they apply. -> references/events-metrics.md ✓ Duration and token metrics carry the required attributes, including gen_ai.token.modality on token counters.
5. Handle versions. Pin the schema URL and, for an existing instrumentation, gate the new conventions behind OTEL_SEMCONV_STABILITY_OPT_IN. -> references/versioning.md ✓ The emitted schema URL and the opt-in behaviour are documented for users.

Verify before done

• Every GenAI span has gen_ai.operation.name with a well-known value where one applies.
• Inference, embeddings, create_agent and client invoke_agent spans have gen_ai.provider.name; server.port is set whenever server.address is.
• Span names follow the patterns above and contain no IDs or prompt text.
• Opt-In content attributes are off unless the user enabled them.
• A failed call sets error.type and ends the span.
• Token counters use {token}, durations use s, and the schema URL is recorded.

Reference index

• references/spans.md: inference, embeddings, retrieval, execute_tool, agent and workflow spans, and MCP spans. Load for step 1.
• references/attributes.md: key gen_ai.* attributes, well-known values, usage attributes, and the content capture patterns. Load for steps 2 and 3.
• references/events-metrics.md: client, server, agent and tool metrics, token metrics, and the GenAI events. Load for step 4.
• references/versioning.md: Development status, the repository move, the schema URL pin, and OTEL_SEMCONV_STABILITY_OPT_IN. Load for step 5.

Related skills

• ocsf to turn AI activity into security audit events: npx skills add ScaleDockHQ/scaledock-skills --skill ocsf.
• cloudevents to carry trace context on events with traceparent: npx skills add ScaleDockHQ/scaledock-skills --skill cloudevents.

Sources

Status uses the publishing body's own maturity term. Checked is the date the source was last read.

• Semantic conventions for generative AI systems: Development, commit b31e9e8 (2026-09-30), Draft posture: build, checked 2026-10-02.
• Registry manifest: development, schema gen-ai-dev/1.42.0-dev on semantic conventions 1.44.0 at commit b31e9e8, checked 2026-10-02.
• Generative client AI spans: Development, commit b31e9e8, checked 2026-10-02.
• GenAI agent and framework spans: Development, commit b31e9e8, checked 2026-10-02.
• Generative AI events: Development, commit b31e9e8, checked 2026-10-02.
• Generative AI exceptions: Development, commit b31e9e8, checked 2026-10-02.
• Generative AI metrics: Development, commit b31e9e8, checked 2026-10-02.
• Generative AI inference token metrics: Development, commit b31e9e8, checked 2026-10-02.
• Model Context Protocol semantic conventions: Development, commit b31e9e8, checked 2026-10-02.
• Gen AI attribute registry: Development, commit b31e9e8, checked 2026-10-02.
• GenAI semantic conventions on opentelemetry.io: Moved, page as served on 2026-10-02, checked 2026-10-02.
• GenAI README with the opt-in transition plan: Development, semantic conventions v1.41.1, checked 2026-10-02.
• Semantic conventions v1.44.0 release: Released, v1.44.0 (2026-08-04), checked 2026-10-02.