Two releases, one date
OpenAI's changelog does not treat November 13 as a single Codex update. It lists 0.168.0 and 0.169.0 as separate releases, each with its own install line, feature list and pull-request trail. The first release brought the side-conversation pane, a provider picker for new chats, audio attachments and animated GIFs. The second, later the same day, extended the pane into multiple tabs, added task search and moved provider configuration into the terminal.
Version 0.169.0 superseded 0.168.0 less than five hours later. Teams evaluating the update should include the later release's trust fix alongside its interface additions. The useful question is not what each version added in isolation. It is what the later release changes for someone who runs the tool on code they did not write.
The fix that outranks the features
The 0.169.0 bug-fix list states the change plainly: codex exec keeps project configuration and MCP servers disabled in untrusted directories, even with --sandbox workspace-write. To load them, a user runs codex exec or codex review with --trust, which explicitly saves project trust.
Project configuration and MCP servers are not inert files. Project config shapes how the agent behaves in that repository, and an MCP server is a program the agent can call. Running an agent inside a freshly cloned repository is a routine move: review this, explain that, run the tests. If the repository's own configuration and servers load in that moment, the code under review gets a say in how the reviewer is configured. The 0.169.0 behavior separates the two decisions: examining a project and trusting it are different acts, and the second now takes a named flag.
The changelog documents a fix keeping project configuration and MCP servers disabled in untrusted directories, including when codex exec uses --sandbox workspace-write. It does not describe the full conditions under which earlier versions behaved incorrectly, and it does not call the old behavior a vulnerability. It lists the change under bug fixes, without an incident, an advisory or an affected-version warning. Readers should take it at that size.
What the headline features actually do
The workspace work is real, and it is gated. In 0.168.0 the side conversation appeared as a single opt-in pane beside the main chat, enabled with features.tui_workspace. In 0.169.0 that workspace supports multiple side-chat tabs, switching by keyboard or mouse, and reopening saved local side chats after a restart. For a developer who parks reference questions beside a long task, that is a genuine workflow gain. For anyone who never enables the flag, nothing changes.
Task search is the quieter improvement. Older tasks can now be searched from the agent command center by name, prompt preview or working directory. That addresses a practical problem in agent tools: the work is easy to start and hard to find again a week later.
Provider setup is the third piece. Custom model providers can be added or edited from the terminal, and with custom gateway onboarding enabled, a provider can be created during sign-in, with a device code for supported gateways over SSH or without a local browser. The release adds an in-terminal setup path, including device-code authentication for supported gateways when custom gateway onboarding is enabled.
The remaining items are smaller. Collapse turn details folds completed activity under each prompt while keeping answers visible and tool output searchable. Ctrl+S saves a diagnostic ZIP locally even when uploads are disabled, which matters in workplaces that forbid sending diagnostics out. The opt-in Cloud connection on macOS and Linux lets the terminal send images, choose a model and reasoning effort, and archive or restore Cloud conversations.
How to update without chasing the version number
Install 0.169.0, not 0.168.0. The official install line is npm install -g @openai/[email protected]. Then check the one behavior that can surprise a script: check automation that expects project configuration or MCP servers to load in an untrusted directory. For a project you intend to trust, the notes document --trust with codex exec or codex review to save that decision explicitly.
Everything else in the release can wait for a reason. The workspace, the Cloud connection and the provider editor are opt-in or on-demand. Teams can evaluate 0.169.0 as their next pinned version after checking configuration-dependent automation, and gain the trust fix without adopting a new interface on the same day.
Sources
- [1] OpenAI, "ChatGPT & Codex changelog" — Codex CLI 0.169.0 entry dated 2026-11-13 (side-chat tabs, task search, provider setup, diagnostic ZIP, Cloud connection, and the codex exec untrusted-directory bug fix quoted in this piece)Read source
- [2] OpenAI, "ChatGPT & Codex changelog" — Codex CLI 0.168.0 entry dated 2026-11-13 (model provider picker, first opt-in features.tui_workspace side pane, audio recording attachments, animated GIF generation)Read source
- [3] OpenAI, Codex CLI 0.168.0 release, GitHub (published 2026-11-13 03:54 UTC)Read source
- [4] OpenAI, Codex CLI 0.169.0 release, GitHub (published 2026-11-13 08:37 UTC)Read source