OpenAI Codex release notes: 0.145.0 imports Cursor and Claude Code work

OpenAI Codex 0.145.0 was published on July 21, 2026 and is now the stable version marked latest on npm. Hammer's previous post covered the alpha release's audio support and multi_agent_v2. The stable release has a more immediate change for teams switching tools: /import can bring supported configuration and work history from Cursor and Claude Code into Codex. Long Codex threads also gain experimental paginated history for efficient resume and search.

Codex CLI is OpenAI's coding agent for the terminal. MCP, or Model Context Protocol, connects an agent to external tools and data sources. Paginated history reads a thread in bounded pages instead of loading everything at once.

OpenAI Codex release notes: what changes in 0.145.0

The GitHub release for 0.145.0 bundles several features and fixes. The npm registry confirms that 0.145.0 is the current stable version, while 0.146.0-alpha.4 is the pre-release at the time of this check.

Audio and multi-agent v2 were already covered in our July 21 post. This article focuses on two other parts of the stable release: moving from Cursor or Claude Code, and the experiment with long, searchable work threads.

Sources: OpenAI Codex 0.145.0 on GitHub. npm registry for @openai/codex.

/import moves more than one settings file

The release notes say /import can now migrate supported settings, MCP servers, plugins, sessions, commands, and project-scoped memories. Cursor support also covers sandbox permissions, project instructions, hooks, agents, and recent chat sessions. If both Cursor and Claude Code contain importable data, the flow asks the user to choose a source.

Plugins have their own migration path. Codex can restore known marketplaces and install enabled plugins from sources including npm, Git, and local directories. A missing or incompatible plugin remains visible in the import result instead of blocking the full installation.

Human step: Start Codex CLI 0.145.0 in the intended project and type /import. Choose a source if Codex detects both Cursor and Claude Code. Read the import result before relying on the migrated configuration. The sources do not promise that passwords or API keys are transferred, so verify MCP authentication and environment variables separately.

Sources: PR #33426, import from Cursor and Claude Code. PR #31672, import of enabled plugins.

Paginated history makes long Codex threads manageable

0.145.0 also adds experimental paginated thread history. The app server can resume a thread with bounded model context and retrieve older records through cursors. The history supports persisted names, sub-agents, and memories.

The experimental search method performs case-insensitive literal search across visible user messages and final agent answers. It returns paginated snippets and a cursor for the matching turn. This is mainly a building block for clients around the Codex app server, not a documented new menu option in the terminal.

Sources: PR #33364, paginated thread history. PR #33907, search in paginated threads.

Short example: use the new Codex feature

Run the prompt after a human has opened the intended project, updated to 0.145.0, and completed /import:

Review the imported Codex configuration against this project. List what came from Cursor or Claude Code, conflicts in project instructions and hooks, MCP connections that lack authentication, and plugins that were skipped. Cite files and import results. Change nothing.

Good output should:

• Separate imported values from the project's existing settings.
• Cite real files and reported import failures.
• Flag missing secrets without printing any secret values.
• Leave the project unchanged and give a short review order.

What teams can test today

Migrate a project that already uses agent instructions, MCP servers, and plugins. Keep the import result as a simple migration receipt: selected source, imported items, skipped items, permissions that need review, and one test showing that Codex calls the intended tool.

The receipt gives an owner something concrete to approve before the project moves over. Tool Forge can help when the same setup needs to connect to live systems with scoped permissions, secrets in environment variables or a secret manager, approval before writes, and retained run logs.