OpenAI's Codex CLI refresh adds parallel agents, but /fork won't isolate your files

OpenAI has shipped what it calls a “major refresh” of Codex CLI, its terminal coding agent, adding a full-screen interface, parallel agent management and session forking. According to a dev.to write-up by YongBo Yu that compiles OpenAI's changelogs and command reference, the refresh is actually a run of builds, 0.156.0 through 0.160.0, shipped between September 22 and October 1, 2026 — so what you see depends on which version you run. The author did not test the build; the descriptions are vendor-reported or drawn from secondary coverage.

What shipped, release by release

Per the compiled changelog:

• 0.156.0 (September 22): an optional full-screen UI via /tui with transcript search, mouse selection and right-click copy; voice input on by default (F8); a /usage analytics dashboard; worktree sessions from the agent command center, with worktree support on by default; Mermaid diagrams and equations rendered in the terminal; and --no-daemon.
• 0.157.0 (September 25): full-screen transcript on by default, automatic background-server startup for eligible sessions, and GPT-6 Sol and Luna added to the catalog.
• 0.159.0 and 0.159.1 (September 29): opt-in instant_interrupt so new input can steer a model mid-response; a compact welcome screen; .aws directories protected by default under writable roots; GPT-6.1 Sol becomes the bundled catalog's default.
• 0.160.0 (October 1): a “show more” history option in the agent command center; projectless sessions where policy permits; queued messages now resume after a reconnect without being sent twice.

A community repost of the announcement also highlights two quality-of-life fixes: older history loads as you scroll past the terminal's own scrollback limit, and the composer stays pinned to the bottom so you can keep typing during long sessions.

A control plane instead of a chat loop

The dev.to post frames the shift as moving Codex CLI from a chat loop to a supervisor for several agent threads. The agent command center — the /agents view in OpenAI's announcement — lists sessions, filters them by status and is where worktree sessions start. Codex Workshop, a secondary source cited in the post, notes that the announcement and the command reference name the subagent commands differently (/agent versus /subagents), so it is worth typing / to see what a given build offers.

Steering is now split across three inputs: Enter injects instructions into the current turn while the agent works, Tab queues a follow-up for the next turn, and pressing Escape twice with an empty composer lets you edit your last message and fork the conversation from that point.

/fork is a chat fork, not a checkout

The sharpest caution in the write-up concerns /fork. Press coverage of the announcement suggested you can fork a conversation “into a worktree with the same context and a separate checkout,” which reads as if /fork always provides filesystem isolation. OpenAI's own command reference is narrower: /fork clones the current chat into a new chat with a fresh ID and leaves the original transcript untouched. A separate checkout comes from a worktree session instead.

The risk is concrete: fork a chat, point both threads at the same directory, and two agents edit one working tree with no isolation. The dev.to author recommends running git worktree list and checking writable roots via /status before a second agent writes anything. Codex Workshop, citing OpenAI's worktree guide, adds that managed worktrees sit under $CODEX_HOME/worktrees on a detached HEAD, that ignored files such as .env are not copied into them, and that Git's one-branch-per-worktree rule still applies.

Practical notes

For a first run, the write-up suggests upgrading with npm install -g @openai/codex@latest, running codex doctor --summary as a health check, and confirming the active model, approval policy and writable roots with /status. Each parallel thread consumes its own tokens, so /usage is where to watch spending. Codex Workshop also warns that approval prompts can come from a thread you are not currently viewing, so check who is asking before approving. If the full-screen UI fights a tmux or SSH setup, /tui sets the UI mode for the next launch and --no-alt-screen disables alternate-screen mode for one run.

Why it matters

The full-screen interface is the visible change, but the structural one is that a single terminal can now supervise multiple agent threads, each with its own context, permissions and possibly its own checkout. That makes “which directory is this agent writing to?” a question developers have to answer deliberately rather than assume. As terminal coding agents turn into orchestrators, the failure mode shifts from one bad suggestion to two agents silently clobbering the same files — and verifying worktrees and writable roots becomes part of the workflow, not an afterthought.