Why it mattersThe worktree flag lets one developer run several Codex sessions against the same repo without stepping on each other's edits, which is the main reason a team using coding agents ends up serialising them behind a single checkout.
OpenAI released version 0.154.0 of its Codex CLI on 9 September 2026 at 22:35 UTC. The release notes list four new features, three of which change how a session runs rather than what the model can do.
Worktrees for forked and new sessions
The change with the widest effect is experimental worktree support. A --worktree flag on session creation, and a /worktree command inside a running session, create an isolated checkout for that session. OpenAI's notes say new and forked sessions can each get their own worktree, and users can browse and resume them later.
A developer running two Codex sessions against the same repository has, until now, either shared one working directory (so the two sessions can overwrite each other's edits) or set up separate clones by hand. This change moves that setup inside the tool. The pull requests referenced are #42652, #43069, #43120 and #43286.
Reply while the agent keeps working
The second change is inline answers. OpenAI describes it as answering questions "while Codex continues working, using suggested choices or custom text, without losing your main draft". A related fix in the same release, #42844 and #42852, says automatic approval reviews now "better preserve authorization context through compaction and reject approvals invalidated by new user instructions".
The stated failure mode being addressed is a session where the user's draft or authorisation is lost when the agent pauses to ask a question. Whether the new flow avoids that in the cases teams hit in practice is the question the release notes cannot answer.
GPT-6 Astra in the model picker
OpenAI says GPT-6 Astra is now available in the model picker and in the Amazon Bedrock catalogues. This is a plumbing change, not a benchmark: the release notes make no performance claim for Astra inside Codex, and OpenAI's own note on the bundled OpenAI Docs skill covers "GPT-6-Astra migration, compatibility, and prompting guidance", so teams switching to it should expect prompt work.
A separate line under Documentation confirms the migration note is in the shipped skill rather than an external post.
Windows and macOS session plumbing
Windows now supports a shared background Codex server, with "daemon lifecycle commands and managed updates" (#42405, #42392). Several other Windows fixes in this release, from sandbox provisioning to control-socket rendezvous, indicate the Windows path is being hardened alongside macOS and Linux rather than left to lag.
On macOS, one line in the bug fixes reads that the sandbox now "blocks terminal input injection" (#42324, #42590). That is a defence against a workspace-controlled helper trying to type into the terminal before Codex has established trust.
Smaller things worth knowing
Vim editing gains an R replace mode with undo and dot-repeat, and escape handling is more reliable in legacy terminals. Copying responses preserves formatting in rich-text apps, and /copy can copy status output or individual session fields. MCP connections now coordinate OAuth token refreshes and surface a login challenge when a refresh fails, without replaying tool calls that had been rejected. The deprecated codex mcp-server entry point is removed in this release, which is the one item a team should check before upgrading.
Source
Source: OpenAI
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.
