Codex 0.165 adds Vertex AI support, but access is still gated

OpenAI released Codex 0.165.0 on October 27, adding a built-in Google Cloud Vertex AI provider and changing the memory default for fresh profiles. The update also improves session navigation and adds GPT-6.1 Terra to the model catalogs. This is a Codex client release, not a new Terra model announcement.

A built-in provider is not an access grant

The Vertex AI implementation recognizes this provider selector:

model_provider = "gcp"

The endpoint is derived from GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION, with global as the default location. The provider uses Google application-default credentials through gcloud by default and permits configuration overrides. That line selects the integration. It is not a complete setup, a credential or proof that the selected project can use a model.

OpenAI’s current Google Cloud access page still describes an invite-only preview for selected customers and supported development and internal testing use cases. Installing a client that understands Vertex does not remove that service-side gate. Our coverage of the Google Cloud partnership explains the separate availability announcement.

The implementation also has a provider-specific capability boundary. It retains function tools and client-side tool search, but disables unsupported features including web search, image generation and reasoning summaries. A workflow that depends on those features should not be assumed to behave identically after changing providers.

Fresh memory profiles get a different default

The memory-default change selects v2 when memory is enabled and the legacy memories directory is absent. Explicit version settings and the dual-write opt-out remain respected. Existing legacy workspaces, and filesystem-check errors, retain the v1 default. This is not a blanket migration of every existing profile.

A separate import change makes selected external-agent memory resources available to the v2 pipeline, while also updating v1 when it is selected or a legacy workspace exists. Import support and default-version selection are related changes, but they are not the same operation.

For an existing user, the practical question is which profile and memory workspace the client is actually using. A clean test profile and a long-running work profile can legitimately behave differently after the same upgrade. Do not erase a working memory directory simply to imitate a fresh-profile default.

Session scope is more explicit

The session-navigation change separates the current directory from the wider repository. The Cwd filter stays local to that directory. A Repo filter appears when multiple local worktrees are available, bringing in sessions from corresponding directories across those worktrees. The agent command center also gains Directory grouping.

Consider two worktrees for a feature branch and a maintenance fix. The repository view is useful when looking for related work across both. The directory view is the narrower choice when you want the session tied to the checkout in front of you. This is an illustrative workflow, not a claim that we tested the release.

Conversation controls and interruption fixes

Reasoning-effort overrides are now enabled by default for supported models, allowing effort changes through conversation configuration updates. Support is model-dependent, so the release does not establish that every provider and model accepts every effort level.

The release also adds experimental attached thread forks to the app-server API. Fixes stop queued agent messages from restarting work after an interrupt or a switch to Plan mode, and stop retries when a provider reports a missing model. Experimental API additions should remain distinct from stable client defaults.

A useful upgrade check is to keep the provider, selected model, profile and working directory visible in your notes. Verify that a resumed session belongs to the intended checkout and that a stopped task stays stopped before entrusting a longer workflow to it. ByteForward reviewed the release notes and implementation descriptions, but did not install or independently test 0.165.0.