Codex CLI 0.164–0.165, Explained: GPT-6.1 Terra in the Catalog, a Vertex AI Provider, /presets, memory v2

After 0.163.0 (2026-10-15) closed out the previous line, openai/codex went quiet for a week — then shipped two stable releases four days apart: 0.164.0 (2026-10-23T04:48 UTC) and 0.165.0 (2026-10-27T20:09 UTC). The first rounds out the interaction and enterprise surfaces; the second writes a cloud provider and GPT-6.1 Terra into the model catalog and flips two defaults. This walkthrough goes through both versions' official release notes item by item (fetched via the GitHub API, checked 2026-10-29), following Codex CLI 0.163.0, Explained: Daybreak On by Default, Instant Interrupts, /config Search.

1. Release timeline and distribution status

On GitHub, the two stable releases sit four days apart: rust-v0.164.0 published 2026-10-23T04:48 UTC, rust-v0.165.0 published 2026-10-27T20:09 UTC (checked 2026-10-29). The 0.165.0 release notes give the cumulative comparison range rust-v0.164.0...rust-v0.165.0, and 0.164.0 gives rust-v0.163.0...rust-v0.164.0 — no overlapping closing fix release this time (nothing like the previous line's 0.162.1).

The distribution surface shows a channel gap worth recording (checked 2026-10-29):

PyPI's openai-codex latest is still 0.163.0 (listed 2026-10-15T21:50 UTC) — both 0.164.0 and 0.165.0 went unlisted, so the same-day three-channel pattern from Codex CLI 0.163.0, Explained: Daybreak On by Default, Instant Interrupts, /config Search did not reappear. The reason for the lag is not stated officially and we won't speculate. The 0.166.0 preview line is already running in parallel: alpha.2 (10-27T11:54 UTC), alpha.3 (10-27T19:23 UTC), alpha.4 (10-28T01:22 UTC), alpha.5 (10-28T04:43 UTC), and alpha.5.1 (10-28T19:17 UTC), all still prereleases as of this check.

2. 0.164.0: /presets, inline /model and /permissions, terminal image previews

Three of the six New Features bullets in 0.164.0 are interaction-side:

• /presets: store reusable prompts and slash commands in ten numbered slots with a keyboard menu for quick access (#54380, #54381) — custom prompt flows previously leaned on config files; there is now an in-session numbered entry point.
• Inline /model and /permissions: set model, reasoning effort, speed, and permission profiles directly with arguments (#54378, #54383) instead of walking through interactive pickers.
• Terminal image previews: on terminals supporting the Kitty or Sixel graphics protocols, the fullscreen transcript can preview viewed and generated images (#54216) — screenshots and generated images no longer require leaving the terminal.

The enterprise-side bullets: enterprise MCP connections can authenticate using eligible Codex account grants or dedicated sign-in, including OAuth clients that require secrets (#55334, #55859, #56001); credentials are masked during supported remote Bash commands and Zsh snapshot sessions, covering custom providers too (#54309, #54432, #56049); the Amazon Bedrock setup wizard is enabled by default, with API key entry grouped alongside other AWS sign-in methods when credentials are detected (#56033, #56023) — Bedrock has carried the GPT-6 Sol/Luna catalogs since 0.157, and the default-on wizard lowers the onboarding friction.

3. 0.165.0: a Vertex AI provider and GPT-6.1 Terra in the catalog

Two catalog-level changes in 0.165.0 deserve their own sections.

First, the built-in Google Cloud Vertex AI provider. The release notes say: "Added a built-in Google Cloud Vertex AI provider, configured with model_provider = "gcp". (#56265)" — Codex now has a CLI-side built-in counterpart to OpenAI's 2026-10-21 announcement of OpenAI models on Google Cloud Vertex AI (see OpenAI models land on Google Cloud: the Responses API on Vertex AI and the SDK vertex() provider). Bedrock followed a similar built-in-provider route earlier (catalogs in 0.157, wizard on by default in 0.164). Beyond the "model_provider takes gcp" sentence, the notes don't expand on auth or region details, and we won't answer for them.

Second, GPT-6.1 Terra enters the model catalog (#57337). We ran a tag-level baseline diff: the models.json on rust-v0.164.0 has no such entry, while the models.json on rust-v0.165.0 adds it — direct evidence that 0.165.0 is the first release to carry it, not an inference from the release notes. Measured fields:

Two caliber reminders: 500K/872K is the Codex-side runtime window (the GPT-5.6 trio sits at 272K/872K in the same file), not an API model-card number; and while GPT-6.1 Terra's API-side model ID and pricing were previously unverified, the identifier gpt-6.1-terra did appear in the SDK type layer on 2026-10-26 (openai-python v3.39.0 #4180 / openai-node v7.45.0 #3054) — see the dated update note at the top of GPT-6.1 Terra announced and live in GitHub Copilot: the everyday agentic coding tier.

4. The two default flips in 0.165.0

• Mid-conversation reasoning-effort changes on by default (#56163): for supported models, switching reasoning effort mid-session no longer requires opting in. 0.163.0 just made instant steering the default; this release does the same for effort adjustment — in-session tuning keeps graduating to defaults release by release.
• memory v2 on by default for fresh profiles (#56211, #56414): newly created profiles default to memory v2, with support for importing external-agent memories into it. Note the wording is scoped to fresh profiles — whether existing profiles migrate is not stated.

Two more catalog-adjacent notes: the session picker gains a repository filter and the agent command center gains directory grouping (#56308), making it easier to navigate sessions across linked worktrees; and the app-server API gains experimental attached thread forks (#56351).

5. Security and stability fixes worth noting

From the two versions' Bug Fixes, the ones relevant to upgrade decisions:

• 0.165.0 closes a marketplace URL validation gap: the release notes say "Closed a marketplace URL validation gap that could allow Git to connect to a different host than the one checked against allowed patterns. (#56263)"; the full changelog titles the PR "Reject backslashes in HTTP(S) marketplace URL userinfo" — read together, the gap involved backslash bypasses in the URL userinfo segment. Deployments using custom marketplaces should prioritize 0.165.0.
• Model-not-found stops retrying (#56445): when a provider reports a model was not found, Codex terminates and preserves the error message instead of spinning on retries.
• Remote MCP tool requests get cancelled (#56132): when their local response futures are dropped, the corresponding remote requests are cancelled, cutting down hung calls.
• 0.164.0 compaction fixes: failed compaction no longer leaves partial summaries in conversation history, and background compaction failures fall back without panicking (#53932, #55581); resuming sessions now finds compressed histories and restores saved thread settings instead of replacing them with global defaults (#54008, #55923).
• 0.164.0 Windows proxy auth: HTTPS and WebSocket connections handle proxy authentication challenges using the signed-in user's Negotiate or NTLM credentials (#55994, #56000).

On the chores side, 0.164.0 does disk hygiene: pruning unused installations, cleaning abandoned marketplace checkouts, and bounding daemon logs (#53827, #55149, #55689, #55690, #55856), plus safety checks for optimized Array.sort operations in the embedded JavaScript runtime (#54519).

6. Queued prompts and preference persistence

One "queued" fix from each version reads well together: 0.164.0 makes queued prompts wait for pending permission selections to resolve before continuing (#55682); 0.165.0 stops queued agent messages from restarting work after an interrupt or a switch to Plan mode (#56330). Together they converge the queuing semantics on "wait for preconditions, don't cross interrupt lines."

Two more 0.165.0 fixes target TUI configuration: fresh TUI sessions honor the server's default approval reviewer when no explicit override is provided (#56090), and TUI preferences stay editable and immediately applicable when execution settings are invalid (#56091).

7. Upgrade guidance

npm users: both @openai/codex and @openai/codex-sdk carry latest = 0.165.0, so upgrade directly; if you want 0.164.0's /presets and inline arguments, 0.165.0 is a superset. PyPI users: openai-codex still sits at 0.163.0 (checked 2026-10-29) — wait for 0.165.0 to be listed rather than mixing installs manually. If you want GPT-6.1 Terra or the Vertex AI provider specifically, 0.165.0 is the minimum — the Terra entry does not exist in 0.164.0's models.json. The 0.166.0 preview line is still in alpha; chasing it means alpha.5.1, but production should wait for stable.

Next steps

• Codex CLI 0.163.0, Explained: Daybreak On by Default, Instant Interrupts, /config Search — the previous line's two default flips
• GPT-6.1 Terra announced and live in GitHub Copilot: the everyday agentic coding tier — Terra's official positioning and availability (with this batch's update note at the top)
• Codex CLI 0.156–0.162, Explained: Voice by Default, GPT-6.1 Sol as Default Model, Worktrees and Command Center — the full walkthrough of the earlier line
• The GPT-6 family grows: gpt-6-sol, gpt-6-luna, gpt-6.1-sol land in the official SDKs — the SDK-layer verification record for the GPT-6.x family