Filing Away the Instructions You Kept Retyping, Into Ten Drawers──Codex 0.164 Lets You Reuse "Stock Phrases" and Swap Models and Permissions on the Fly

On October 23, 2026, OpenAI released version 0.164.0 of "Codex CLI," its terminal-based coding agent. The highlight isn't a new model but quiet refinements that cut down the repetitive steps in day-to-day use. Tuck away the instructions you kept retyping into a "drawer," swap models and permissions on the spot, and tighten how credentials are handled──it's a set of updates that pay off the most for people who run the agent for long stretches.

Putting the Instructions You Kept Retyping Into a "Drawer"

The biggest addition is /presets. You can save frequently used prompts and slash commands into ten numbered slots and call them up from a keyboard menu. The settings are stored in the user's configuration file (tui.presets), and a picker, an editor, and a confirmation screen have been provided.

Stock instructions like "In this repo, write the tests first" or "Review in this order" no longer have to be typed out from scratch every time. It's effectively a way to pull your own "stock phrases"──previously buried in chat-style exchanges──out into reusable components.

Swapping Just for the Moment, Without Changing Your Defaults

You can now also drive session controls directly. By passing arguments to /model and /permissions, you can switch the model, reasoning effort, speed, and permission profile in a single move, without wading through menus. This isn't a permanent change to your defaults; it's a temporary setting that applies to the session at hand.

The aim is to keep "what settings am I running under right now" from becoming ambiguous. A deep model for heavy investigation, a fast model for light formatting, pared-down permissions in someone else's repo──you can declare this kind of switching without breaking your workflow.

Tightening How Credentials Are Handled

Defensive measures were added as well. Here are the two main points.

• Credential masking: In supported remote Bash commands and Zsh snapshot sessions, credentials are concealed. Credentials originating from custom providers are also covered.
• Enterprise MCP authentication: Internal MCP connections can now be authenticated via a Codex account grant or a dedicated sign-in. OAuth clients that require a secret are supported too. Account-level grants are stored in the OS keyring and are automatically revoked when the account or configuration changes.

Alongside this, fixes were included so that a failed compaction no longer corrupts the conversation history, resuming a session correctly restores the compacted history and thread settings, and the Amazon Bedrock setup wizard is enabled by default.

The Homework That Remains Behind the Convenience

That said, it would be premature to take this update as a guarantee that misconfigurations or leaks disappear. Commentary on the release points out the following caveats.

• Masking isn't a cure-all: Remote credential masking retains a fallback that, in execution environments able to handle only built-in providers, "runs without concealing" while issuing a warning. It doesn't cover every path.
• Presets go stale: Saved stock instructions drift out of sync with reality as a repo's conventions change. Behind the convenience comes the operational chore of "continually updating old stock phrases."
• Credentials that change mid-session: Enterprise authentication tidied up the entry point, but operational questions remain──such as what happens if credentials change over the course of a long-running session.

These points aren't backed by published figures demonstrating their effect. If you adopt it, the sound approach is to verify in your own environment "how it actually behaves amid your usual interruptions and setting changes"──rather than the promotional feature list──before rolling it out more widely.

What Changes

The thrust of 0.164 lies not in the model's smarts but in shaving off "the friction of running the agent as an everyday tool". Turning stock phrases into components, making setting changes explicit, and reducing credential exposure──all of these are improvements premised on long, repeated, team-based use rather than on one-and-done short exchanges. You could call it one more step in the trend where the axis of competition among coding agents is broadening from "one-shot brilliance" to "operational robustness."

References: openai/codex release 0.164.0 / Remio: presets and session controls in Codex 0.164.0 / Developers Digest: /presets and inline /model and /permissions / AI-TLDR: Codex CLI 0.164.0