OpenAI Codex rust-v0.154.0 stable was tagged 2026-09-09 22:37 UTC, ~24 hours after Tuesday's Charter Pillar on Claude Code v2.1.265+266. v0.154.0 supersedes the v0.154.0-alpha.1 through alpha.11 nightly churn that the desk assessed NOVELTY_LOW in earlier sweeps. The release adds five new capabilities that materially affect how teams run Codex, plus the GPT-6-Astra rollout into Codex itself.
From the v0.154.0 release notes (fetched 2026-09-10 14:08 UTC), verbatim:
--worktree or /worktree, then browse and resume them." PRs #42652, #43069, #43120, #43286. Each worktree is an isolated copy of the project; sessions in one worktree don't affect files in another. The flag is currently labeled experimental.R replace mode with undo and dot-repeat, plus more reliable Escape handling in legacy terminals. PRs #42194, #42584. Replaces the existing character-level r with line-level R. Dot-repeat (.) repeats the previous R action./copy can copy status output or individual session fields. PRs #42847, #43055.Worktrees are the largest practical change. Multi-agent workflows that fork a project to make a parallel change without colliding with the main worktree are now a first-class Codex primitive. The flag is experimental, so behavior may evolve; the underlying capability — isolated checkouts that don't share filesystem state — is real.
Inline questions close a long-standing Codex friction: when Codex would ask for clarification, the user's draft was lost. The new primitive keeps the draft in place, prompts the user with suggested choices or free text, and resumes Codex on the answer. This is operationally significant for long-running agent workflows where the user's draft matters more than the answer's timing.
Windows server sharing closes the gap that Windows users previously had to start a new Codex per terminal. The daemon-lifecycle commands and managed updates mirror what Linux and macOS users have had since earlier Codex versions; this brings Windows to parity for shared-agent workflows.
Bedrock catalogs mean regulated or AWS-pinned teams can adopt GPT-6 Astra without leaving their existing VPC/identity boundary. This is a deployment-cost and compliance-posture improvement for any team that previously had to dual-route through both OpenAI and Bedrock.
Vim replace mode is small but real: the gap between Codex Vim and real Vim has narrowed. Teams that standard on Vim in Codex won't lose muscle memory on R.
All claims sourced verbatim from openai/codex/releases/tag/rust-v0.154.0 (fetched 2026-09-10 14:08 UTC). Cross-referenced against the GitHub atom feed (openai/codex/releases.atom, updated 2026-09-10T11:06:18Z) for release timestamp. Each PR referenced in the release notes is linked in the bullet list above; the article does not re-fetch each PR.
No firsthand test run was performed against the Codex binary or a Bedrock endpoint. Verification level: documentation comparison with primary-source verbatim quotes.
--worktree means behavior may evolve without deprecation notice. Teams adopting worktree workflows should pin the Codex version and document which experimental flags are in use.v0.154.0 is a substantive stable release. Five new capabilities (worktree, inline questions, Windows server sharing, Vim R replace, copy formatting) plus the GPT-6 Astra rollout into Bedrock. Three of the five (worktree, inline questions, Windows server) are first-class for multi-agent workflows; the other two are quality-of-life. None are breaking changes for existing workflows. The experimental label on worktrees is the only caveat worth flagging.
rust-v0.154.0. The Bedrock catalog addition alone is worth the upgrade for any AWS-pinned team.--worktree on a non-production workflow to confirm the isolated-checkout behavior matches your expectations. The experimental label means the API may evolve; flag any reliance on the flag in your documentation.Originally published: 2026-09-10 14:08 UTC Last verified: 2026-09-10 14:08 UTC No corrections at this time.