Codex CLI 0.165.0 Release Notes: Vertex AI Provider, Memory v2 Default, Reasoning Effort Switching

Codex CLI 0.165.0 was tagged on October 27 as rust-v0.165.0. The release notes list six new features and six bug fixes. The ones most likely to touch your setup: a built-in Google Cloud Vertex AI provider, memory v2 as the default for fresh profiles, and reasoning effort changes during a conversation.

This article is based only on the GitHub release page. I did not read the Codex documentation, the pull requests, or the source, and I did not run the release. Everything below is "the release notes say", not a tested result.

What the release notes list as new

1. A built-in Vertex AI provider. The notes say it is configured with model_provider = "gcp" (PR #56265). The changelog also has test entries for Vertex auth fixtures (#56293).

2. GPT-6.1 Terra in the model catalogs (#57337). The changelog labels this a backport to the 0.165.0 alpha line.

3. Session navigation. A repository filter in the session picker and directory grouping in the agent command center, aimed at sessions across linked worktrees (#56308).

4. Experimental attached thread forks in the app-server API (#56351).

5. Reasoning effort changes mid-conversation, now enabled by default for supported models (#56163).

6. Memory v2 on by default for fresh profiles, plus support for importing external-agent memories into it (#56211, #56414).

The changelog also lists a few related memory entries, including experimental thread memory summary reads (#56465) and a fix to stop memory imports from racing with consolidation (#56462). The release notes do not explain what "memory v2" stores or how it differs from v1, so I will not guess.

What the release notes list as fixed

• Fresh TUI sessions now honor the server's default approval reviewer when no explicit override is given (#56090).
• TUI preferences stay editable and apply immediately when execution settings are invalid (#56091).
• Remote MCP tool requests are cancelled when their local response futures are dropped (#56132).
• Queued agent messages no longer restart work after an interrupt or a switch to Plan mode (#56330).
• Retrying stops when a provider reports that a model was not found, and the provider's error message is kept (#56445).
• A marketplace URL validation gap is closed. The notes say it could allow Git to connect to a different host than the one checked against allowed patterns (#56263). The matching changelog entry says backslashes in HTTP(S) marketplace URL userinfo are now rejected.

The changelog has more entries than I list here (169 pull request lines on the page). I read the headline sections and searched the rest for the items above. Other changes, such as AWS SigV4 signing for network proxy hooks (#56182, #56277), appear there without a user-facing explanation.

Why the Vertex provider is the item to watch

A built-in provider means a Codex setup can point at a Google Cloud model endpoint with one config value, instead of a custom provider block. That is my reading of the line item. The notes do not say which models are available through it, how it authenticates, or how billing works. Those details decide whether it is useful for you, and they are not on the release page.

What I would check

These are my untested suggestions.

1. Read the config docs for gcp before enabling it. Find the auth method and the supported models, and check who pays.

2. Back up your profile before you upgrade if you rely on memory. The default change applies to fresh profiles, per the notes. If you have an existing profile, confirm which version it uses.

3. Test reasoning effort switching on one model. The notes say "supported models" only.

4. If you use marketplaces, check your allowed URL patterns. The fix rejects some URL forms, so an unusual entry might stop working.

5. If you depend on interrupt or Plan mode behavior, rerun a short session that uses both after you upgrade.

6. Watch the 0.166.0 alphas. The releases page lists 0.166.0-alpha.5 as a pre-release dated October 28. I did not read its notes.

What I would not conclude

• That the Vertex provider supports every model you use.
• That memory v2 is better, safer or cheaper. The notes make no such claim.
• That the security fix was exploited. The notes describe a validation gap and say nothing about abuse.
• That 0.165.0 is stable for your workload. I did not run it.

Frequently asked questions

When was Codex CLI 0.165.0 released?

The release page shows it as rust-v0.165.0, dated October 27, with commit 360319d.

How do I use the Vertex AI provider?

The notes say to set model_provider = "gcp". They give no further setup steps, so read the docs.

Does 0.165.0 add GPT-6.1 Terra?

The notes say GPT-6.1 Terra was added to the model catalogs.

Is memory v2 on for everyone?

The notes say it is on by default for fresh profiles. They do not say what happens to existing profiles.

Can I change reasoning effort mid-conversation?

The notes say it is enabled by default for supported models. They do not list which models are supported.

What security fix is in this release?

A marketplace URL validation gap. Per the notes, it could let Git connect to a different host than the one checked against allowed patterns.

Is there a newer release?

The releases page also shows 0.166.0 alpha builds, the latest being alpha.5 dated October 28. Those are pre-releases.

Sources

All read on October 28, 2026.

• Releases, openai/codex (GitHub): the 0.165.0 notes, changelog entries, and the 0.166.0 alpha listing.

Not read: the Codex docs, the pull requests, the source, and the 0.166.0 alpha notes.