The maintainer, 2026-10-18: "please work overnight with codex-cloud-2 to satisfy full conceptual parity with trunk, keeping to simple, clear layers with absolutely minimal code and data driven readability."

Trunk, the ocx branch at a46d73f5bf, unchanged since parity-v1.md was written, was read against ocx2. About 20 gaps were found, 11 of them not in parity-v1.md: questions put to a person and what waits on them; conversation carried across turns; writing conventions given to each session; an agent reading its own thread; an agent asking for a build and learning how it ended; reading verdicts; seeing what a running session is doing; a session ending itself; a workspace kept between runs; trunk's CI and build templates; and comparing agents on one task. The rest were already listed: run/codex and a tried run/claude; a long session a person joins; CI per check-in; unjudged check-ins; a list of runs; removing a finished run's workspace; the native agent.

Each gap got a design, data first (a pattern or a run's keys) before any code; then the designs were checked against the code and ordered into 22 check-ins, about 940 lines of source on version 1. The audit also found four timing faults in today's code, since fixed on agent-tools (6d5ac4cf8c, f3138e4316).

The attachment parity-audit.tar, plain, holds the inventory, every design, and the order.