agents · review · first seen Oct 28, 09:33 UTC

OpenAI adds Vertex AI provider to Codex 0.165.0

Codex 0.165.0 shipped a built-in Google Cloud Vertex AI provider configured with model_provider = "gcp". For a builder, the reason to look now is that Vertex AI is a named provider option in Codex; the release notes do not establish how it will perform or what it will cost in your workflow.

L0 · primary1 source4 claims checked4 min read

What shipped

The central change is the built-in Google Cloud Vertex AI provider. The release notes give one configuration detail: model_provider = "gcp". I would treat that as the starting point for an evaluation, not as a complete setup guide. The supplied source does not specify credentials, supported models through this provider, or any provider-specific limits. [1]

The same release added GPT-6.1 Terra to the model catalogs. That is a catalog change, not evidence that Terra is available through the new Vertex AI provider. Keep those two additions separate when deciding what to test: the notes establish that each was added, but do not state a relationship between them. [1][2]

For app users, the release notes also describe a repository filter in the session picker and directory grouping in the agent command center for sessions across linked worktrees. The app-server API gained experimental attached thread forks. These are distinct changes: navigation affects how a user finds sessions, while attached forks are an experimental API feature. Neither is stated as a requirement for using the Vertex AI provider. [1][2]

Memory v2 is enabled by default for fresh profiles, and the release notes say it supports importing external-agent memories. The fresh-profile qualification matters. I would not assume an existing profile changes its memory behavior on update, because that is not what the supplied text says. [1]

Why it matters if you build

If your team has been deciding whether Codex can be configured with Vertex AI as a built-in provider, this release gives you a concrete option to inspect today: model_provider = "gcp". It does not, by itself, answer whether your preferred model, authentication arrangement, or operating constraints work with that option. Make the decision from a trial in your own environment rather than from the provider name alone. [1]

The session-navigation changes are relevant to builders who move among linked worktrees. A repository filter and directory grouping give them specific places to look when finding a session. The notes do not quantify time saved or describe every navigation behavior, so I would evaluate the picker and command center against the way my team actually organizes work. [1]

Two changes deserve separate test plans. Attached thread forks are explicitly experimental, so an app-server integration should not treat their presence as a stability commitment. Memory v2 defaults to on for fresh profiles, so a clean-profile check may reveal behavior that an existing-profile check does not. The supplied notes give no prices, quotas, performance measurements, or migration procedure; those remain open questions before a broader rollout. [1][2][3]

How to try it today

1. Check that the Codex release you are evaluating is 0.165.0 before attributing any observed behavior to these additions.
2. Inspect the built-in provider configuration and try model_provider = "gcp" in a test setup. Record what additional setup your environment requires; it is not specified here.
3. Check the model catalog for GPT-6.1 Terra, but test its availability separately from the Vertex AI provider.
4. If you use linked worktrees, open the session picker and agent command center to assess the repository filter and directory grouping.
5. If you use the app-server API, evaluate attached thread forks as an experimental feature, separate from production assumptions.
6. Compare a fresh profile with your existing profile before drawing conclusions about the memory v2 default.

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. shipped L0 · primary

   Codex 0.165.0 adds a built-in Google Cloud Vertex AI provider configured with model_provider = "gcp".

   per GitHub releases · openai/codex (L0)
2. shipped L0 · primary

   Codex 0.165.0 adds GPT-6.1 Terra to the model catalogs.

   per GitHub releases · openai/codex (L0)
3. shipped L0 · primary

   Codex 0.165.0 adds experimental attached thread forks to the app-server API.

   per GitHub releases · openai/codex (L0)
4. shipped L0 · primary

   Codex 0.165.0 enables memory v2 by default for fresh profiles.

   per GitHub releases · openai/codex (L0)

What is not confirmed

The supplied release excerpt does not establish Vertex AI authentication steps, supported models through that provider, pricing, quotas, or performance. It also does not say that GPT-6.1 Terra is available through Vertex AI. Those are checks to make, not facts to infer from two additions appearing in one release. [1][2]

The notes do not establish a stability commitment for experimental attached thread forks or say that existing profiles enable memory v2 by default. They do not quantify the effect of the session-navigation changes. [1][2]

Who should act, who should wait

Act now if a built-in Vertex AI provider is the missing option you need to evaluate, or if sessions across linked worktrees are a routine part of your work. Keep the first pass limited to a test setup and verify the unresolved setup and model questions there. Wait on a production decision that depends on known costs, limits, model availability through the provider, or stable attached forks: the supplied notes do not settle those points. [1][2][3]

Sources

• 0.165.0 — GitHub releases · openai/codex L0 · primary