OpenAI Shows Codex CLI with Voice and Git Worktrees
OpenAI demonstrated Codex CLI running voice commands and isolating concurrent agents across Git worktrees to stop local repo file collisions.
Video walkthrough details concurrent agents and physical Git folders designed to eliminate local file collisions.
OpenAI published a technical demonstration on its official @OpenAIDevs account on X showing how the Codex CLI handles software engineering workflows. The three-minute video shows developers triggering tasks by voice, tracking simultaneous runs from a consolidated view, and separating local branches using Git worktrees.
Distributed through the public npm registry under @openai/codex, the CLI brings together features shipped across September updates and presented during DevDay 2026. The setup targets a familiar hurdle in terminal-based agent systems: file contention when several tasks attempt to edit the same local repository tree simultaneously.
Inside the Codex CLI terminal workflow
In the demonstration, developers open the audio channel by pressing F8 or typing /voice. The model transcribes spoken instructions and launches the corresponding agent task, eliminating the need to type out lengthy prompts during debugging or refactoring passes.
Managing these concurrent tasks happens inside the /agents view. The screen lists the status of each running process, lets users pause active jobs, displays recent action history, and allows switching between code branches within a single terminal window instead of juggling separate terminal tabs.
Branch isolation with Git worktrees
To prevent parallel agents from clobbering identical files on disk, the Codex CLI integrates directly with Git worktrees through the /fork and /worktree commands. The native Git capability creates separate physical directories linked to the exact same local repository.
This separation allows one agent to generate unit tests on a feature branch while a second agent patches a bug on another branch, without either run interfering with uncommitted changes in the main working tree. OpenAI’s technical documentation does not include automated merge conflict handling: reviewing diffs and executing the final merge remains a manual developer responsibility.
According to release logs tracked by OpenTools, experimental worktree support debuted in CLI release 0.154.0 on September 9. Voice input followed in version 0.155.0 on September 17, and both capabilities stabilized alongside the /agents dashboard in release 0.156.0 on September 22.
Pricing rules, practical limits, and UI tools
Running tasks in the Codex CLI draws from standard message and agent quotas included with paid ChatGPT and Codex subscriptions. Streaming audio follows a different policy: accounts on flexible billing are charged separately for active voice connection time, according to OpenAI support documentation. OpenAI has not published official limits for continuous voice session lengths or latency benchmarks for remote SSH sessions.
Beyond agent orchestration and voice processing, the terminal environment renders Mermaid diagrams, formats mathematical formulas with LaTeX, includes a /usage command for checking credit balances, and provides collapsible diff blocks to inspect changes before writing them to disk.