OpenAI Codex release notes: CLI 0.138.0 hands threads to Desktop

OpenAI has a new Codex changelog entry: Codex CLI 0.138.0, published on June 8, 2026. The useful part is not another model headline. It is a cleaner handoff between the terminal and Codex Desktop. For teams that already let a coding agent read files, propose changes, and ask for approval, that can make the work easier to follow.

Source: OpenAI Codex changelog, Codex CLI 0.138.0

OpenAI Codex release notes: what changed in CLI 0.138.0

Codex CLI is OpenAI Codex in the terminal: a local coding agent that can inspect a project, reason about changes, and run tools within the permissions you set. The 0.138.0 release notes include a few changes worth testing now:

• /app can hand off the current CLI thread into Codex Desktop on macOS and native Windows. Windows workspaces can also open directly into Desktop instead of stopping at a manual prompt.
• Local image attachments and standalone image generations now expose saved file paths to the model, making follow-up edits and file references more reliable.
• Reasoning effort selection is more flexible in the TUI, including fallback shortcuts for terminals where Alt bindings do not work.
• App-server integrations can read account token usage, and Codex auth supports v2 personal access tokens in CLI and app-server flows.
• Plugin automation has richer structured output: add/remove and marketplace commands support --json, plugin lists include marketplace source, and plugin details can expose default prompts, remote MCP servers, and unavailable app templates.

Source: GitHub release rust-v0.138.0 and npm @openai/codex 0.138.0

Why this matters in practical automation work

A coding agent is an AI assistant that can work inside a code project, not just answer in a chat box. A sandbox limits what the agent may read or change. An approval gate is a pause where a human must approve a step before it runs.

What matters in 0.138.0 is that Codex behaves a little more like a workbench. The terminal can be where you start narrow, with explicit commands and controls. Desktop can be where you review the thread, images, diffs, and next steps more visually. For Nordic teams building internal AI workflows, that is often where the value is: not that the agent acts with less supervision, but that the handoff becomes easier to inspect.

If Codex connects to real systems, keep the controls boring and clear: environment variables instead of secrets in chat, a secret manager where needed, scoped API keys, least privilege, redacted logs, and approvals before sensitive changes. That is Tool Forge work: design the tool flow so the agent can help without making access vague.

Human step before the prompt

These are human steps, not instructions for the agent:

• Upgrade Codex CLI if you want to test the verified version:

npm install -g @openai/[email protected]

• Start or continue a local Codex CLI thread in a project you can review.
• On macOS or native Windows, use /app when you want to hand off the current CLI thread into Codex Desktop.

Source: OpenAI Codex changelog, 2026-06-08

Short example: use the new Codex feature

Paste this into Codex before using the /app handoff:

Prepare this thread for Desktop handoff with /app. Review the current diff, open risks, and next visual step. Write a short handoff note with files, tests, and what I should inspect in Desktop. Do not edit files.

Good output should include:

• which files and decisions already belong to the thread
• which tests or checks have run, and which are still missing
• what is safe to review visually in Desktop
• where a human should approve before the agent continues

What we are watching next

0.138.0 also includes bug fixes that make Goal workflows, TUI editing, config errors, and AGENTS.md loading more predictable. That sounds less exciting than a new button. In practice, these fixes are often what decide whether an agent workflow feels safe enough to use several times a week.

Source: GitHub release rust-v0.138.0