OpenAI ships a wave of Codex agent updates: web search, Desktop handoff, mobile worktrees

OpenAI pushed a cluster of updates to its Codex agentic coding tools on June 8 and 9, 2026 — spanning the terminal CLI, the desktop app, and the mobile experience inside ChatGPT. None of it is a headline new model, but taken together the releases sharpen Codex as a day-to-day coding agent, and several changes matter for anyone wiring AI assistants into a real workflow.

What landed in the CLI

The biggest moves are in the Codex CLI, which shipped versions 0.138.0 (June 8) and 0.139.0 (June 9) according to the project's GitHub releases.

Version 0.139.0 adds standalone web search in code mode, returning plaintext results the agent can read directly — useful for checking a library's current API without leaving the terminal. It also makes the CLI a better Model Context Protocol citizen: tool schemas now preserve oneOf and allOf JSON-schema structures, which previously got flattened and could confuse MCP-exposed tools. The release also improves the codex doctor diagnostic to report your editor and pager environment, and fixes resume --last and fork --last argument handling.

Version 0.138.0 introduced a /app command that hands a session off to Codex Desktop on macOS and Windows, so you can start in the terminal and continue in the GUI. It also adds support for v2 personal access tokens, exposes saved file paths for local image attachments to the model, and improves discovery of AGENTS.md instruction files in remote and symlinked workspaces.

Codex on your phone

On the mobile side, OpenAI's Codex changelog lists a June 9 ChatGPT for iOS release (build 1.2026.153) that brings genuinely agentic features to the phone: branch selection, worktree creation, a dedicated Codex profile screen with usage statistics, and /goal support for managing objectives on the go. That follows a June 4 desktop app update (26.602) adding activity insights and shareable profile cards.

Why it matters for builders

The theme across these releases is less about raw model capability and more about making an AI agent fit into the way developers actually work — multiple surfaces (terminal, desktop, mobile), cleaner tool interoperability through MCP, and parallel work via git worktrees. Worktree support on mobile is a small but telling detail: it assumes you might kick off an isolated agent run from your phone and review it later, a pattern that mirrors features in competing tools.

If you use Codex, the practical takeaways are to update the CLI to pick up the MCP schema fix if you rely on custom tools, and to try the /app handoff if you bounce between terminal and GUI. If you're tool-shopping, this is another data point that the major AI coding assistants are converging on the same agentic primitives — background tasks, worktrees, MCP, and project-level instruction files — rather than competing purely on model benchmarks.

As always, treat changelogs as the source of truth: feature flags and rollout timing can differ from release notes, so check codex --version and your update channel before assuming a given capability is live.