Emit GenAI message events only on the chat span, with v1.36.0 bodies

[shibayan]

Problem

With sensitive-data capture enabled, the per-message GenAI telemetry events (gen_ai.system.message / gen_ai.user.message / gen_ai.assistant.message / gen_ai.tool.message / gen_ai.choice) were emitted on both the invoke_agent span and the chat span, so every exchange was double-reported. The reference implementations emit these events for the model invocation only:

• Python ([BREAKING] Python: Consolidate OTel GenAI Semantic Conventions versions microsoft/agent-framework#7673, released in python-1.15.0) split message capture into event emission and span-attribute capture, and removed event emission from the agent telemetry layer — invoke_agent spans now carry attribute-form content only.
• .NET's OpenTelemetryAgent has never emitted message content itself; it delegates to the chat-client instrumentation.

The event payload was also off-spec: it carried a role+parts JSON serialization (the shape of the gen_ai.input.messages / gen_ai.output.messages span attributes) under the v1.36.0 event names — a hybrid belonging to neither semconv generation.

Change

• setMessageContent now records span attributes only. The gen_ai.input.messages / gen_ai.output.messages attributes are unchanged on both span kinds.
• A new addMessageEvents (internal to core) emits the events on the chat span with the OTel GenAI v1.36.0 structured bodies, matching the Python emitter:
  • system / user messages: {"content": text}
  • assistant messages: {"content"?, "tool_calls"?} with tool_calls entries {id, type: "function", function: {name, arguments}} (only calls with both a call id and a name)
  • tool messages: one event per function_result with a call id, body {"id", "content"} (null/undefined results become "")
  • responses: one gen_ai.choice per message, body {"index", "finish_reason", "message": {content?, role?, tool_calls?}}; role is included only when it is not assistant
• A response without a finish reason emits no choice events. The finish reason falls back to the raw representation's finish_reason when the normalized field is absent, matching Python's _get_response_finish_reason.
• Every event carries gen_ai.system (the provider name) as an attribute, and instructions are emitted ahead of the input messages as a gen_ai.system.message event.
• Timestamps advance 1 microsecond per event from a single wall-clock read, so ordering survives backends that collapse tight timestamps. The base uses performance.timeOrigin + performance.now(): Date.now() has millisecond resolution, so the input and choice batches of one fast invocation would collide and the choice events would stamp earlier than the stepped input events.
• Events remain span events (the OpenTelemetry logs API lives outside @opentelemetry/api, the core's only runtime dependency); the structured body rides the body attribute as JSON, and event.name is kept for backends that lift span events into log records.

Not ported (upstream additions that are opt-in switches rather than fixes): the ENABLE_MESSAGE_EVENTS gate and the OTEL_SEMCONV_STABILITY_OPT_IN semconv-generation switch.

Breaking

Consumers reading message events from invoke_agent spans, or parsing the previous content attribute payload, need to read the chat span's events and the JSON body attribute instead. Span attributes are unaffected.

Verification

• Reproduction tests were written first and observed failing against the previous behavior (agent-span events present, old payload shape, choice emitted without a finish reason, no provider attribute).
• After the fix, the implementation was temporarily reverted to confirm the same tests fail again, then restored.
• pnpm check passes (exit code verified).

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR aligns core OpenTelemetry GenAI message telemetry with the reference implementations by emitting per-message GenAI events only on the chat span (avoiding double-reporting on both invoke_agent and chat) and by switching event payloads to the OTel GenAI v1.36.0 structured bodies (JSON stored in the body attribute), while keeping the existing span-attribute message capture unchanged.

Changes:

• Split message capture into (a) span-attribute recording (setMessageContent) and (b) per-message event emission (addMessageEvents) scoped to chat spans only.
• Implement v1.36.0 event shaping (including provider stamping and timestamp stepping) and add responseFinishReason fallback logic.
• Update/extend tests to lock in span placement, event shapes, provider attribute stamping, timestamp ordering, and finish-reason behavior.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
packages/core/src/observability/tracing.ts	Removes per-message events from attribute capture; adds v1.36.0 message/choice event emitter, timestamp stepping, and finish-reason fallback helper.
packages/core/src/observability/tracing.test.ts	Adds focused tests asserting correct span placement and exact v1.36.0 event bodies, provider stamping, timestamp ordering, and finish reason fallback.
packages/core/src/observability/metrics.test.ts	Updates expectations to reflect events being emitted on chat spans and body JSON usage.
packages/core/src/observability/attributes.ts	Adds the v1.36.0 provider attribute key (gen_ai.system) to the GEN_AI constants.
packages/core/src/client/telemetry.ts	Wires addMessageEvents into chat-span telemetry for both input and output sides, using responseFinishReason.

Suppressed comments (2)

packages/core/src/observability/tracing.ts:155

• For gen_ai.tool.message, the v1.36.0 body shape expects content to be a string. Currently content.result is unknown, so the event may emit non-string content (object/array) or even become fully [unserializable] for values like bigint/circular structures. Coercing the tool result to a non-throwing string keeps the event schema stable and preserves the call id even when the raw result isn't JSON-serializable.

    for (const content of message.contents) {
      if (content.type === 'function_result' && content.callId !== '') {
        events.push({
          name: GEN_AI_MESSAGE_EVENT.tool,
          body: { id: content.callId, content: content.result ?? '' },

packages/core/src/observability/tracing.ts:176

• For gen_ai.system.message / gen_ai.user.message events, the v1.36.0 body shape is always { "content": text }. Currently content is omitted when textOfContents(...) is empty, producing {} for non-text-only messages (e.g. image-only user input), which changes the event schema and makes it harder for consumers to rely on a stable shape. Consider always emitting content for non-assistant roles (empty string when there's no text), while keeping assistant content optional.

  const body: Record<string, unknown> = {};
  const text = textOfContents(message.contents);
  if (text !== '') {
    body.content = text;
  }

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.