Give Your Ever-Growing Threads a "Name" and a "Pin": How Codex 0.146 Turns Conversations into a Tidy Workspace and Starts Sharing a Plugin Shelf with Claude Code

Version v0.146.0 of OpenAI's coding agent "Codex" was released on July 29, 2026. This update isn't about a flashy new model; instead, it centers on the "art of tidying up your work" and "carrying your extensions with you"—improvements that matter most to people who run agents day in and day out. It also takes a step toward sharing a plugin shelf with other tools.

Tidying up open-ended conversations with names and pins

Use an agent long enough and conversations (threads) pile up by purpose—research, implementation, review—until you can no longer tell which was which. Version 0.146.0 rolls in a set of tools to sort out this clutter.

• Name your sessions: You can give a name to a new session started with /new or /clear, making it easier to tell them apart in a list later.
• Pin important threads: You can pin the main thread you're working on so it doesn't get buried under a pile of history.
• Side conversations you can switch to without closing: You can open a separate discussion as a "side conversation" running in parallel, without closing the thread you're working on.
• Fork your history: You can branch a thread while carrying over its existing history. You can also create "temporary forks" that don't show up in the list, so you can run trial experiments without polluting the main line.

Each of these is unremarkable on its own, but the aim is consistent. The direction is to let you treat a dialogue with the agent not as a "disposable, one-off" but as a workspace that you can name, pin, and branch to manage. The more you hand long-running, multi-task work to an agent, the more this organizational cost translates directly into real productivity.

Sharing the plugin shelf across tool boundaries

The other pillar is interoperability around extensions. Version 0.146.0 adds the following support.

• Agent Plugins manifest support: Codex can now read the common format (manifest) that defines a plugin.
• Publishing plugins within a workspace: You can distribute and share plugins inside your own workspace.
• Expanded marketplaces: As sources for obtaining plugins, support has been added for the Amazon Bedrock and Claude Code marketplaces.

The third point is worth watching. The design lets Codex pull in extensions even from the plugin shelf prepared for Claude Code—supposedly a competitor. This is less a one-off feature addition than an extension of an industry-wide trend. The month before, Codex 0.145 made it possible to migrate settings, MCP servers, and sessions wholesale from Cursor or Claude Code via /import, and GitHub Copilot CLI adopted the common plugin format (Open Plugin Spec). Each company is converging toward a state where "an extension you build once can be reused with a different agent."

From a user's perspective, this is a change in the direction of loosening lock-in to any particular tool. Even if you switch agents, it becomes easier to carry over the assets—the skills and plugins—you've built up.

Laying the groundwork to get through corporate networks

Improvements that matter in real operations also made it in. The app-server can now connect to a remote Code Mode host over WebSocket, adding a path that links a remote execution environment with a local client. Alongside this, the release has been fixed to consistently honor a configured proxy across authentication, plugin downloads, MCP authorization, remote execution, WebSocket, redirects, and connections to LM Studio. On corporate networks where external communication is only possible through a proxy, this guarantee that "every connection goes through the proxy" becomes a low-key but crucial factor that decides whether adoption is even feasible. There are also signs of consideration for in-house and on-prem-leaning setups, such as enabling standalone web search even with custom model providers.

How to read this as a business signal

This update shows that coding agents are maturing from the stage of "fighting over the smartest model" into tools that teams can run reliably. Being able to organize conversations with names and pins lowers the practical, on-the-ground cost of running agents across multiple people and multiple tasks. Plugin interoperability moves tool selection from a "once you decide, you're locked in" bet toward a "you can move it later" choice, lowering the psychological cost of switching.

At the same time, the more entry points there are for pulling in extensions from outside, the bigger the operational challenge becomes of who verifies the trustworthiness of what you're pulling in, and how. Some point out that third-party plugins obtained via marketplaces trade convenience for a heightened need for measures such as verifying the supplier and isolating them in a sandbox. In fact, companies are simultaneously moving to narrow the scope of execution—recently, Copilot CLI shifted toward tightening its sandbox by default. If you want to take advantage of how portable extensions have become, the realistic approach is to prepare the governance design—which plugins to allow, and where to run them in isolation—with the same intensity you devote to convenience.

In summary, 0.146.0 is a bundle of individually small improvements, but it's an update that assembles another layer of the underpinnings needed for real operations: "managing your work with an agent, carrying your assets with you, and running it even on corporate networks."

References: Codex rust-v0.146.0 release notes (GitHub) / Codex changelog (ChatGPT Learn, official) / Codex Updates by OpenAI (Releasebot) / Open Plugin Spec (vercel-labs)