OpenAI Codex release notes: CLI 0.135.0 sharpens diagnostics and profiles

This is the kind of OpenAI Codex release-notes update that looks small until you think about operations. Codex CLI 0.135.0 is not a big demo. It is about seeing what the agent is running inside, which profiles apply, and whether the remote session is actually the one you think it is.

Codex CLI is the command-line tool for OpenAI Codex, a coding agent that can read a repository, propose changes and run checks inside the boundaries you set. A sandbox is the bounded environment where the agent's commands can run. An approval gate is the point where a human must approve an action before the agent continues.

OpenAI Codex release notes: what changed in CLI 0.135.0

OpenAI published Codex CLI 0.135.0 on May 28, 2026. The official changelog mainly points to better diagnostics and clearer control inside the terminal interface, TUI.

• codex doctor reports richer environment, Git, terminal, app-server and thread inventory diagnostics for support cases.
• /status shows remote connection details and server version when the TUI is connected over a remote transport.
• /permissions now understands named permission profiles and displays configured custom profiles.
• Vim mode gained text-object editing, better word and line-end behavior, and a configurable interrupt-turn binding.
• The Python SDK exposes friendlier Sandbox presets for thread and turn APIs.

Source: OpenAI Codex changelog: Codex CLI 0.135.0 and GitHub release: rust-v0.135.0

Why it matters for controlled agent work

This is useful for teams that have moved past "can Codex write code?" and are now asking "can we review Codex as part of our operating model?" Diagnostics and profiles are not glamorous, but they decide whether an agent environment can be debugged when something gets weird.

For Swedish and Nordic organizations, this matters when Codex is connected to a repository, MCP servers, internal tools or remote environments. MCP, Model Context Protocol, is a way to connect an agent to external tools and data sources. Permissions need to be visible: which tools are read-only, which can write, which sandbox applies, and where should a human approve the next step?

This belongs in Tool Forge work. Build an integration people can trust: use environment variables or a secret manager for keys, scoped permissions, separate profiles for test and production, redaction of sensitive fields where needed, approval gates before write actions and logs someone can actually follow.

What you can test today

Human step: upgrade Codex CLI and run diagnostics in a repo where Codex already does work. The official changelog shows the installation command for this version, and the release note describes the expanded codex doctor report.

npm install -g @openai/[email protected]
codex doctor

In the TUI, you can then check /permissions. If the session runs over remote transport, /status can show remote connection details and server version.

Source: OpenAI Codex changelog: installation command, codex doctor, /status and /permissions

Short example: use the new Codex feature

Human step: run the commands above, copy the relevant part of codex doctor, and inspect /permissions before you start a larger agent task. Then paste a short instruction into Codex:

Use this Codex 0.135.0 diagnostics snapshot and permission-profile output to review whether this repo is ready for a controlled agent session. Identify environment issues, remote-session facts, active permission profile, sandbox or approval risks, and the smallest safe next task. Do not modify files.

Good output should:

• separate environment issues, permissions, and the code task itself
• name the active permission profile and what it allows
• point out whether remote status or server version is missing when it should exist
• suggest one small next step without changing files

What to watch next

There are already alpha tags after 0.135.0, but their public release text does not give the same useful decision support. For most teams, the stable signal right now is simple: make the Codex environment easier to review before you ask the agent to do more work.