If your company runs its models through a cloud account rather than a direct vendor login, a CLI agent has to meet you there. This week Codex adds Google Cloud Vertex AI as a built-in provider, so Vertex joins Amazon Bedrock in the catalog story. Around it sit smaller changes that make long sessions less fragile: reasoning effort you can change mid-conversation, memory v2 as the default, and a transcript you can select without touching the mouse.
This update covers Week 44, October 27-29, 2026: versions 0.165.0 and 0.166.0, two releases with 28 changelog entries. The Vertex provider is the headline. The rest is mostly session handling and TUI work.
Vertex AI as a provider: what you get and what to check
It is built in. Codex now ships a Google Cloud Vertex AI provider. You enable it with one line in your configuration: model_provider = 'gcp'. The provider arrived in 0.165.0, and 0.166.0 extended its model catalog.
The catalog gets GPT-6.1 Terra. In 0.166.0, GPT-6.1 Terra was added to the bundled, Vertex AI and Amazon Bedrock model catalogs. If your organization routes traffic through Google Cloud or AWS, you can pick the same model on either route without going through a custom provider definition.
What it is good for. A built-in provider means you no longer maintain your own provider block for Vertex. For teams whose billing, access control and data-residency rules already live in Google Cloud, that removes one piece of hand-written config from every developer machine.
The conditions that matter, as far as the changelog states them:
- The provider is selected with
model_provider = 'gcp'. - The Vertex AI catalog includes GPT-6.1 Terra from 0.166.0.
- The same model is in the bundled and Amazon Bedrock catalogs as of 0.166.0.
What remains open. The changelog entries say nothing about authentication, project selection, regions or plan requirements. Check the Codex documentation for the setup details before you roll this out to a team. (A one-line config switch rarely covers the whole story on the cloud side.)
Changes for everyday work
GPT-6.1 Terra is in the catalogs. It first appeared in the model catalogs in 0.165.0. With 0.166.0 it reached the bundled, Vertex AI and Bedrock lists.
Reasoning effort can change mid-conversation. For supported models, this is now enabled by default as of 0.165.0. You can lower effort for a routine edit and raise it when the problem turns hard, without starting a new session.
Memory v2 is on by default for fresh profiles. It also supports importing memories from external agents. If you keep notes in another agent, the import path means you do not need to retype them. The default applies to fresh profiles, so check the setting on an older profile rather than assume it matches.
You can select transcript text from the keyboard. 0.166.0 adds keyboard-driven selection in the TUI transcript, with word, line and page motions, integrated search and configurable shortcuts. The changelog does not list the default key bindings, so look at the shortcut configuration in the documentation.
codex queue --resume reaches into older threads. It queues a message and then resumes the thread if it is unloaded. You can push a follow-up into an older session in one step instead of resuming it first.
Technical notes
Sessions and threads.
- The session picker gained a repository filter, and the agent command center groups by directory. Both make it easier to find your way across linked worktrees (0.165.0).
- Queued agent messages no longer restart work after an interrupt or a switch to Plan mode (0.165.0).
- Acknowledged approval and permission settings are now preserved when you resume or fork a thread (0.166.0). You should no longer have to confirm the same grants twice.
- Reasoning-token accounting is preserved across resume and fork, which avoids premature compaction (0.166.0).
TUI and diagnostics.
codex doctor --network-onlyruns focused configuration, connectivity and WebSocket diagnostics (0.166.0). It fits a proxy or firewall problem better than the full doctor run.- The TUI now gives clearer explanations when a request is blocked for safety reasons, and it offers selectable safer follow-up prompts (0.166.0).
Providers and errors.
- Codex no longer retries when a provider reports model-not-found, and the error message is preserved (0.165.0). With a new provider and new catalog entries, a wrong model name now fails fast and shows you the provider’s own message.
App-server API.
- Experimental attached thread forks were added to the app-server API (0.165.0). The word experimental applies, so do not build anything stable on it yet.
What to take away
If you work in the terminal:
- Try the keyboard transcript selection and check the shortcut configuration for the bindings.
- Change reasoning effort mid-conversation on a supported model and see whether it fits your routine edits.
- Look at what memory v2 holds on a fresh profile, and try the external-agent import if you keep memories elsewhere.
If you run parallel sessions:
- Use the repository filter in the session picker and the directory grouping in the agent command center if you juggle linked worktrees.
- Try
codex queue --resumeto feed a follow-up into an unloaded thread. - Resume or fork a thread with approvals already granted and confirm that the settings carry over. Long resumed sessions should also compact later than before.
If you administer integrations or company deployments:
- Test
model_provider = 'gcp'on one machine and read the documentation for the authentication and project requirements, which the changelog does not cover. - Check that GPT-6.1 Terra shows up in the Vertex AI and Bedrock catalogs in your setup.
- Use
codex doctor --network-onlyas the first step when a developer reports connectivity or WebSocket trouble. - Treat the attached thread forks in the app-server API as experimental.
Sources: the Codex CLI changelog, versions 0.165.0 to 0.166.0 (October 27-29, 2026), and the Codex documentation. The attached thread forks in the app-server API are experimental, and the availability of beta features may change.
