Folding the "Conversation" Back into a "Workshop"—How Codex 0.156 Made git worktree the Default and Turned Terminal Agents into a Work Desk You Can Run in Parallel and Resume

OpenAI's coding agent Codex has, through a series of updates in late September, shifted its character from "a disposable conversation" to "a workshop you can fold up and carry with you." The centerpiece is 0.156.0, released on September 23. Here, "worktree sessions" backed by git worktree became enabled by default, and alongside it landed a full-screen /tui mode, a /usage dashboard for viewing usage status, voice interaction on by default, a /daemon for operating a locally resident server, and the rendering of Mermaid diagrams and math within responses. The following 0.157.0 (September 25) added the new GPT-6 Sol / Luna models and their Bedrock support, while 0.158.0 (September 28) and 0.159.0 were devoted to fixing finer issues such as cloud sync for large workspaces and session survivability.

No longer a "one-off chat"

The most practically significant change this time is the seemingly modest move of making worktree sessions the default. Sessions opened with --worktree or /worktree ride on top of the worktree feature Git has had all along (a mechanism for lining up separate checkouts per task without cloning the entire repository). On a foundation already familiar to developers, the agent's work becomes tied to an independent file state.

To borrow the phrasing of an explainer article, agent sessions are no longer "disposable chats" but "durable workshops equipped with isolation, resumption, permissions, plugin state, and review boundaries." The meaning of this shift is easier to grasp when broken into the following elements.

• Isolation: You can begin risky changes or trial-and-error in a separate worktree and proceed without affecting the main branch.
• Resumption: Even if you pause partway through, the session survives and you can pick up where you left off. While waiting on a decision, reversible work can still move forward.
• Permissions: Authorization state is preserved across resumptions and forks, and approval reviews now retain context more easily even through summarization (compaction).
• Plugin state: Tools auto-update even within a running session, and OAuth token handling makes authentication less likely to expire during long-running executions.
• Review boundaries: You can inspect the diff and session state as an "artifact," then selectively merge it or discard the entire checkout.

"How much you've used" and "the work screen" were tidied up at the same time

0.156.0 also touches on ease of work. The /usage dashboard lets you trace your account's usage status, total token volume, and the activity of plugins and skills. In the agent command center, you can now filter tasks by state and spin up worktree sessions from there. The full-screen /tui mode supports transcript search and mouse operation, making it easier to follow long work logs. As a finer point, voice interaction is now on by default and can be toggled with F8.

Leaning toward a parallelism-first workflow

Put together, you can see a trend of terminal agents shifting from a tool that "watches over a single conversation to the end" to a work desk built on the premise of "running multiple pieces of work simultaneously and bundling them later." By splitting a worktree per task, you can run exploratory refactoring and your main implementation in parallel and check each diff independently. Because pausing and resuming work, the work itself is not lost even if you're stalled waiting on a decision. The visibility provided by /usage serves as a clue for retroactively measuring the consumption that tends to balloon under this kind of parallel operation.

From a business standpoint, the ability to conduct reviews against "the diff as an artifact" fits well with how implementation teams operate. You can receive agent output on a per-branch basis, pick only the good changes to incorporate, and throw away the misses along with the entire checkout—handling unattended, semi-automated work in a flow close to the existing Git review culture.

The reins that remain behind "on by default"

On the other hand, caveats behind the convenience have been pointed out as well. The explainer article warns that "worktrees can still get messy, and if scope is loose the agent may reach for the wrong files, too." The point is that discipline is required to get the benefits. Specifically, it lists clearly delimiting the scope of work, treating permissions as part of the review, and recording tool changes alongside the task's artifacts. The very fact that worktree sessions became enabled by default may itself surface as a behavioral change for users who assumed the previous experience, so it's safest to revisit your scope and permission design before building this into unattended operation.

Behind the flashy new-model buzz (GPT-6 Sol / Luna), Codex's late-September updates are quietly rebuilding the design of where agents "work." From disposable conversations to isolated, resumable workshops. This change in the underlying foundation may prove to have a longer-lasting effect on day-to-day development workflows.

References: Codex Updates by OpenAI - Releasebot / Codex CLI Worktrees Turn Agent Runs Into Durable Sessions - Developers Digest / Codex CLI Changelog - Gradually / Releases · openai/codex / OpenAI Codex CLI Releases, Summarized - Havoptic