OpenAI Codex PR #25089, “Compress cold local rollouts,” was merged into main on June 1, 2026. The PR describes readers that handle compressed rollouts and materialize them back to plain .jsonl before append writes. Its merge description said the background compression worker was then under development behind the disabled-by-default local_thread_store_compression flag and initially handled only archived_sessions. This repository merge does not establish the local client build, flag state, or why a particular file was compressed.
The local sample was recorded at 2026-10-14 11:37:01 UTC (19:37:01 in Shanghai). The check streamed 16 .jsonl.zst files through a decoder to count bytes; it did not parse or retain JSONL session text. Per-file counts were not preserved.
**I byte-counted 16 local .jsonl.zst files by streaming them through a decoder; I did not parse or retain session text. The compressed files total 129,491,499 bytes and the decoded output totals 214,184,564 bytes. The 84,693,065-byte difference is 39.54% of the decoded-byte total.**
These numbers describe only the 16 files. They are not a Codex-wide compression rate and do not identify when compression occurred or the local feature configuration.
A related community issue, #34337, was opened by fengjikui on July 20, 2026. Its current body reports a separate 760,247,640-byte rollout becoming 463,794,252 bytes through the author-described level-3/default-window path, about 1.64x. That is the issue author's separate measurement, not an OpenAI product guarantee or a reproduction of this local 16-file check.
Codex repository history also lists a September 1, 2026 commit titled “Include shared histories in rollout compression.” That is a source-history entry; it does not identify the installed build or feature flag on this machine.