Add response and conversation IDs to completed-response OTLP usage spans

What variant of Codex are you using?

CLI.

What feature would you like to see?

Include two existing IDs on completed-response spans that contain token usage:

• gen_ai.response.id: the ID supplied by the provider in the completed response.
• gen_ai.conversation.id: the existing Codex ThreadId.

We collect Codex usage through OTLP. These attributes would let us associate usage with a provider response and group usage by thread across turns and resume. Codex already has both IDs, but the usage spans do not export them.

Example of the requested span attributes, shown as a decoded map:

{
  "gen_ai.usage.input_tokens": 30,
  "gen_ai.usage.output_tokens": 7,
  "gen_ai.response.id": "resp-example",
  "gen_ai.conversation.id": "<existing Codex ThreadId>"
}

Expected behavior:

• Export the response ID with the usage from the same completion.
• Omit the response ID attribute when the ID is empty. Do not generate a replacement.
• Keep the conversation ID consistent across turns and resume.
• Use different conversation IDs for separate threads.
• Preserve token values, existing event fields, parser behavior, and exporter settings.
• Add no metric dimensions.

Additional information

The relevant code is the handle_responses span declaration and SessionTelemetry::record_responses. I also checked the completion-recording code at upstream commit 5c5308fc9a9ee789049d646ef11e5400384b9c6f; it still omits both attributes.

A reference implementation and OTLP export test are available in my fork (download patch). The patch applies to 0f7c8608bdce4874acba3f3a4467f1bf8275fecc. Please use or adapt it if helpful. This is a feature request, in line with the contribution policy; no upstream PR is open.

The test sends synthetic responses through Codex and checks the actual OTLP HTTP payload. It covers repeated response IDs, different response IDs with equal usage, empty IDs, existing missing-ID parser behavior, multiple turns, resume, and separate threads.

just test -p codex-core --test all completed_response_identity_reaches_otlp_exporter

Related: #43078 asks about correlation for reasoning-effort completion logs. This request is limited to two existing IDs on completed-response usage spans.

Would the team consider implementing this behavior? We can provide further reproduction details.

Start with the handle_responses span declaration in codex-rs/core/src/session/turn.rs and SessionTelemetry::record_responses in codex-rs/otel/src/events/session_telemetry.rs. Review the linked reference implementation and run just test -p codex-core --test all completed_response_identity_reaches_otlp_exporter. Done means completed-response usage spans export the existing response and thread IDs with the specified empty-ID, resume, and separate-thread behavior while preserving existing telemetry.