[hooks] Expose per-model-request timing in rollout for observability integrations

Feature request

Please expose stable timing for each model sampling request in the persisted rollout available through the Stop hook's transcript_path, or provide an equivalent model lifecycle hook.

For each sampling request, observability integrations need:

• model request start time
• first output time, or the existing per-request ttft_ms
• model response completion time
• a stable correlation key linking those timings to emitted response_item records, token usage, and the existing turn_id

This would allow integrations to map the lifecycle to Langfuse generation.startTime, generation.completionStartTime, and generation.endTime without inferring model timing from tool events.

Current behavior

The current HookEventName enum contains these actual variants:

PreToolUse, PermissionRequest, PostToolUse, PreCompact, PostCompact, SessionStart, SessionEnd, UserPromptSubmit, SubagentStart, SubagentStop, and Stop.

There is no model-request, response-start, first-output, or response-completed hook. Source: HookEventName.

The current StopRequest contains these actual fields:

session_id, turn_id, cwd, transcript_path, model, permission_mode, stop_hook_active, last_assistant_message, and target.

Source: StopRequest.

A real Codex rollout contains a turn-level terminal record with this field shape (values redacted; field names are unchanged):

{
  "timestamp": "<redacted>",
  "type": "event_msg",
  "payload": {
    "type": "task_complete",
    "turn_id": "<redacted>",
    "started_at": "<redacted>",
    "completed_at": "<redacted>",
    "duration_ms": "<redacted>",
    "time_to_first_token_ms": "<redacted>"
  }
}

The public protocol defines the same actual fields on TurnCompleteEvent, and documents time_to_first_token_ms as the duration between turn start and the first model token. Source: TurnCompleteEvent.

A turn can contain multiple sampling requests separated by tools, so this turn-scoped value cannot identify completion start for later generations.

The relevant per-request rollout boundary currently has this actual field shape (values redacted; field names are unchanged):

{
  "timestamp": "<redacted>",
  "type": "response_item",
  "payload": {
    "type": "custom_tool_call",
    "id": "<redacted>",
    "status": "completed",
    "call_id": "<redacted>",
    "name": "exec",
    "internal_chat_message_metadata_passthrough": {
      "turn_id": "<redacted>"
    }
  }
}

{
  "timestamp": "<redacted>",
  "type": "event_msg",
  "payload": {
    "type": "token_count",
    "info": {
      "last_token_usage": {
        "input_tokens": "<redacted>",
        "cached_input_tokens": "<redacted>",
        "cache_write_input_tokens": "<redacted>",
        "output_tokens": "<redacted>",
        "reasoning_output_tokens": "<redacted>",
        "total_tokens": "<redacted>"
      }
    }
  }
}

These records do not expose the enclosing sampling request's start, first output, or completion timestamps. A post-hoc tracing integration therefore has to infer generation boundaries from response_item and token_count; in tool-calling turns that can make generation timing collapse onto, or include, tool execution.

Existing signals that could be reused

Codex already computes relevant timing internally:

• TurnCompleteEvent.time_to_first_token_ms is exposed, but it is turn-scoped.
• PR #30883 attaches the actual per-request field ttft_ms to the existing codex.sse_event / response.completed telemetry record.
• ItemStartedEvent.started_at_ms and ItemCompletedEvent.started_at_ms / completed_at_ms describe item lifecycles, not the enclosing model sampling request. Source: ItemStartedEvent and ItemCompletedEvent.

The requested change could reuse the existing per-request ttft_ms measurement rather than measure TTFT again. Please persist/expose it together with model-request start and response-completion timing at the hook/rollout boundary. The exact new schema is intentionally left to maintainers; all field names shown above are fields Codex already emits or defines today.

Reproduction

1. Configure a Stop hook that reads transcript_path.
2. Run a turn with multiple model requests separated by tool calls.
3. Inspect the rollout around response_item and event_msg.token_count.
4. Attempt to reconstruct one generation observation per model request.
5. The turn-level time_to_first_token_ms cannot be assigned to later requests, and no per-request model lifecycle is available.

Related

• #30883 added per-request TTFT completion telemetry.
• #33420 requests a streamed assistant-message hook, but assistant text deltas do not cover reasoning-only or tool-call responses and do not provide request start.
• #37317 discusses native OTel response signals and their volume.
• #35363 added item start times, which helps item timing but does not establish the model-request lifecycle.

pip install codex-rescue==0.1.0a5
codex-rescue doctor --latest