Supported metadata-only correlation for reasoning-effort completion telemetry

In Codex CLI 0.153.4 (GPT-6 Astra, ChatGPT authentication), an isolated app-server experiment submitted one ephemeral synthetic turn with explicit effort low while the isolated profile default was high. The normal turn completed successfully. Supported JSON OTLP export produced two codex.sse_event / response.completed log records for that conversation: one without model_reasoning_effort and one with model_reasoning_effort=low. Neither completion record carried trace/span, turn, request or response IDs. Separate spans carried the turn ID. No raw events or private interfaces were enabled.

The public rust-v0.153.4 source copies the constructed request's effort into request-scoped telemetry and emits it at completion; the completion function itself receives usage and timing but not request/turn identifiers.

Could you identify a supported way to:

1. Associate every completion event with its exact turn, request/response and retry attempt?
2. Distinguish connection warmup from actual inference without guessing from token counts?
3. Confirm the semantics of model_reasoning_effort: constructed client request, server-accepted setting, or effective server execution setting?
4. Observe the relevant identity/semantics while retaining only allowlisted metadata and keeping prompts, account identifiers and raw tool results out of logs?

The public rollout-trace README describes CODEX_ROLLOUT_TRACE_ROOT and captures the required inference identities, but record_started/record_completed also persist full request/response payloads. Is there a supported metadata-only or pre-write-redacted mode that records these identities and the request effort without storing full content?

Relevant version-pinned source:

• https://github.com/openai/codex/blob/3d2ee51ca2d5db578f328aa75e20aa22c0197c9a/codex-rs/rollout-trace/README.md
• https://github.com/openai/codex/blob/3d2ee51ca2d5db578f328aa75e20aa22c0197c9a/codex-rs/rollout-trace/src/inference.rs
• https://github.com/openai/codex/blob/3d2ee51ca2d5db578f328aa75e20aa22c0197c9a/codex-rs/otel/src/events/session_telemetry.rs#L1011

We will keep automated effort selection disabled unless these semantics and correlations can be verified. We are not requesting private IPC, TLS interception, internal raw events, or access to an existing desktop task.

Read the version-pinned rollout-trace README and codex-rs/rollout-trace/src/inference.rs, then inspect codex-rs/otel/src/events/session_telemetry.rs around line 1011. Reproduce the described supported JSON OTLP experiment and determine whether the requested identities and reasoning-effort semantics can be observed without full payloads. Done means a supported metadata-only path or a documented conclusion about its absence.