Field report: no schema_version in rollout files — silent vocabulary change makes third-party analytics read 0 turns for every session

Summary

We run a fail-closed session-analytics layer over Codex rollout files (715 sessions/7 days across 3 production repos, multi-vendor orchestration). Between CLI 0.151.x and 0.153.4 the turn-carrying record changed from event_msg/user_message to response_item/message (with input_text/output_text content parts) — with no schema_version marker at any level and no deprecation event.

Our v1 parser (pinned to the old vocabulary) read every newer session as 0 turns, silently. We now handle both shapes and fail loudly on zero-recognition, but the marker would have made the transition detectable in minutes instead of days.

Reproduction

A minimal, fully tokenized repro (two rollout files + a reference old-vocabulary parser + expected/actual table) is here:
https://github.com/oguzhanpisgin/ai-session-analytics/tree/main/benchmarks/codex-silent-zero

A structural specimen of all observed record types (values tokenized) is here:
https://github.com/oguzhanpisgin/ai-session-analytics/tree/main/specimens/codex

Requests

1. schema_version marker per rollout file (or per line), bumped when the record vocabulary changes.
2. Structured task_role / verdict fields for rollouts: in our data, 130 of 364 sessions were verification tasks where a substantiated BLOCK verdict is the successful outcome. Naive end-of-session error scoring reported 61% success; role-aware scoring reports ~94% (paired counts and intervals in the linked repo). First-class fields would remove the need for text inference entirely.

Notes

• JSON key order is not semantic and we do not rely on it; the request is about an explicit version discriminator.
• A separate security report will be filed through the coordinated disclosure channel for an unrelated observation (an opaque credential-shaped string in a reasoning record; no live credential matched).

Start by reviewing the linked benchmarks/codex-silent-zero reproduction and specimens/codex, then locate the rollout record definitions and serialization paths in the repository. Clarify the versioning and task-role requirements with maintainers before changing them. Done means the requested metadata is emitted and the existing rollout shapes remain distinguishable to consumers.