OpenAI Codex release notes: 0.146.0 names sessions and adds Agent Plugins

OpenAI released Codex CLI 0.146.0 on July 29, 2026. A new model is not the headline this time. OpenAI instead cleaned up the work itself with named sessions, side conversations that stay open, and portable Agent Plugins. Developers and automation teams can separate parallel work without losing track of the tools and instructions attached to it.

OpenAI Codex 0.146.0 release notes: what changed?

Codex CLI is OpenAI's local coding agent for the terminal. A coding agent can read a codebase, propose changes, and use tools within the permissions granted by its user.

Name the main thread and keep side questions open

/new and /clear now accept an optional session name. Codex can also persist pinned thread state and let clients filter for pinned threads. When a side conversation is open, the default ctrl-/ binding switches between it and the parent thread without closing either. The footer shows the active binding, and the binding is configurable.

The release can also fork threads that use paginated history. Temporary forks can stay out of the normal thread list. That is useful when an implementation plan needs to remain intact while a bounded question about test data or an API contract gets a separate review.

Source: OpenAI Codex 0.146.0 release notes.

Source: OpenAI PR #34605 on session names.

Source: OpenAI PR #35011 on side conversations.

Pinning relies on an app-server capability. The release does not document a dedicated pinning slash command, so the visible control depends on the Codex client you use.

Source: OpenAI PR #34840 on persisted thread pinning.

Agent Plugins gain a portable manifest

An Agent Plugin is a package that describes reusable agent capabilities in a manifest. Codex 0.146.0 recognizes Agent Plugins 1.0 through a root plugin.json. Codex can map portable metadata, direct skills under skills/, and MCP configuration in mcp.json. MCP, the Model Context Protocol, is a protocol for connecting an AI agent to external tools and data sources.

Codex-specific apps, hooks, and interface settings can live in the com.openai extension. The app server can also tell a client whether it may publish a plugin to the workspace directory. The client should fail closed when that capability is unavailable. OpenAI also lists marketplaces for Amazon Bedrock and Claude Code in the stable release.

This makes a plugin easier to treat as a managed asset rather than local special configuration. Keep the manifest in version control, store API keys in environment variables or a secret manager, and give each MCP server scoped permissions. Writing tools need clear approval gates and run logs.

Source: OpenAI PR #35105 on Agent Plugins 1.0 manifests.

Source: OpenAI PR #35254 on workspace plugin publishing.

Proxy and MCP handling become less brittle

Version 0.146.0 honors configured proxies across authentication, plugin downloads, MCP authorization, remote execution, and WebSockets. MCP connections can be reused when configuration remains unchanged, while closed connections are replaced and authentication changes trigger a refresh. That is a practical operations improvement for companies behind a proxy or running several tool servers, even though sessions and plugins are the clearest user-facing changes.

Source: OpenAI Codex 0.146.0 release notes, bug fixes.

What this changes for Nordic teams

A named main session can hold the implementation plan. Side conversations can handle questions that need separate reasoning without closing or muddling the main thread. Agent Plugins give the team a clearer home for reusable skills and tool connections.

This is useful in client projects where developers, integration specialists, and reviewers work in the same repository. The name makes the work searchable, the plugin manifest shows which capabilities are included, and the run log shows what the agent actually did.

Human step

1. Update Codex CLI with the official npm command:

npm install -g @openai/codex

1. Start Codex with codex.
2. Create a named session with a command such as /new customer-portal-integration. To clear the current session and start again with a name, use /clear customer-portal-integration.
3. When a side conversation is already open, ctrl-/ switches between it and the parent thread by default. Check the footer if the binding has been changed.

Source: OpenAI Codex README with the installation command.

Source: OpenAI PR #34605 with optional session-name syntax.

Source: OpenAI PR #35011 with the default side-conversation binding.

Short example: use the new Codex feature

Replace the bracketed text and send this prompt in the named session:

Review this repository for [the change]. Return one main implementation plan and a short list of questions that belong in side conversations. Include affected files, tests, dependencies, and which enabled plugin or skill each step relies on. Change nothing.

Good output should:

• cite real files in the repository,
• separate the main plan from bounded side questions,
• explain why each plugin or skill is needed,
• stop before file edits, so a person can approve the next step.

Test one workflow today

Choose an existing integration task with a clear main objective and at least one question that usually interrupts the work. Run the mapping in a named session. Handle the side question in an available side conversation, then confirm that you can return to the main plan without losing context. If the project uses plugins, compare the agent's list of active skills and MCP connections with the manifests that are actually present in the workspace.

GitHub and npm now list 0.146.0 as the stable version. That differs from the July 27 article, when 0.145.0 remained stable and 0.146 was available only as an alpha.

Source: npm metadata for @openai/codex 0.146.0.

Teams that want to make named agent workflows and portable plugins reviewable can take the next step through Tool Forge: choose one workspace, one plugin manifest, and one approval gate, then run the setup against a real integration need.