Codex CLI 0.166.0, Explained: GPT-6.1 Terra in Vertex AI and Bedrock Catalogs, Transcript Search, doctor
0.166.0 went stable on 2026-10-29T19:47 UTC: GPT-6.1 Terra covers the Vertex AI and Bedrock catalogs too; the TUI gains keyboard transcript selection and search; queue --resume and doctor --network-only land. PyPI gap: three versions.
openai/codex published 0.166.0 stable on 2026-10-29T19:47 UTC (checked via the GitHub API on 2026-10-30), closing out the six-step alpha preview line that ran from Oct 27 to Oct 29 (alpha.2 through alpha.6, all prereleases) — the stable release and the final alpha share a day. It lands just two days after 0.165.0 (Oct 27, 20:09 UTC). Both npm packages listed it within nine minutes; PyPI's gap widened to three versions. Everything below is checked against the release notes text.
1. Release cadence and distribution
| Channel | Status (checked 2026-10-30) |
|---|---|
| GitHub release | 0.166.0 stable, 2026-10-29T19:47:53 UTC; preview line alpha.2 (Oct 27, 11:54) through alpha.6 (Oct 29, 01:41) |
| npm @openai/codex | latest 0.166.0, listed 2026-10-29T19:56:19 UTC |
| npm @openai/codex-sdk | latest 0.166.0, listed 2026-10-29T19:56:16 UTC |
| PyPI openai-codex | latest still 0.163.0 (Oct 15, 21:50 UTC); 0.164.0 / 0.165.0 / 0.166.0 all missing |
The official cumulative comparison range is rust-v0.165.0...rust-v0.166.0. The PyPI gap has grown from two versions (0.164, 0.165) to three; the reason is not stated officially.
2. GPT-6.1 Terra: from one catalog to three
The third item in the New Features section: Added GPT-6.1 Terra to the bundled, Vertex AI, and Amazon Bedrock model catalogs (#57293).
There is a detail worth unpacking, and we tested it at the tag level: the bundled models.json (codex-rs/models-manager/models.json) on rust-v0.166.0 is byte-identical (same md5) to the file on rust-v0.165.0. Terra's bundled-catalog entry therefore already existed in 0.165.0 (see our 0.164–0.165 line explainer), and what 0.166.0 actually adds is the Vertex AI and Amazon Bedrock catalog coverage. The Terra entry is identical in both tags: visibility=list (selector-visible), priority 7, context_window 500000 / max_context_window 872000, default effort low — and, as always, the 500K figure is the Codex-side runtime window, not an API model-card number; API-side context and pricing remain unverified on our end as of 2026-10-30.
For context, the same models.json still ranks GPT-6.1 Sol (priority 1) and GPT-6-Astra (priority 2) at the top, with GPT-6.1 Terra holding position 7.
3. Interaction surface: transcript keyboard selection, unified Markdown, safety-block explanations
- Keyboard-driven transcript selection: word, line, and page motions, integrated search, and configurable shortcuts (#57221, #57222, #57223, #57224). This is the TUI transcript's first systematic keyboard navigation.
- Consistent Markdown rendering: async questions, choices, answers, and transcript replies all render Markdown (#57259, #57292).
- Safety-block experience: clearer safety-block explanations in the TUI, plus selectable safer follow-up prompts (#57363, #57365).
4. Two command surfaces: queue --resume and doctor --network-only
codex queue --resume resumes unloaded threads after a message is queued (#57232) — it completes the queued-prompt capability introduced in the 0.164 line with a recovery path for which thread the queued message belongs to.
codex doctor --network-only runs focused configuration, connectivity, and WebSocket diagnostics (#57606). The doctor command itself is not new; the flag that scopes it to network-side checks is.
5. Reliability fixes: the resume/fork path is the focus
| Fix | What it does | PR |
|---|---|---|
| Approvals and permissions | resume/fork preserves acknowledged approval and permission settings | #57227 |
| Reasoning tokens | Preserves reasoning-token accounting across resume/fork to avoid premature compaction | #57839 |
| Composer restore | Keeps local image attachments when restoring killed composer text | #57502 |
| Descendant threads | Restores spawned descendant threads when unarchiving the parent | #57635 |
| app-server threads | Fixes thread startup/shutdown races; releases completed runtimes promptly | #57787, #57788 |
| TUI reconnection | Improved reconnection recovery, including offline /quit and /exit handling | #57820 |
| macOS cleanup | Fixes standalone-release cleanup for in-use installs and non-ASCII paths | #57503 |
6. Under the hood and release engineering
Three chores that are invisible to end users but shape distribution: Bazel switched to a source-built V8 with target-matched Rust bindings and native Windows MSVC archive support (#57164, #57269, #57644, #57832); the standalone executor now ships in the canonical CLI and app-server release packages (#57837); development and release builds are standardized on Node.js 24.22.1 and Rust 1.95.0 (#57168, #57373).
7. Notable PR titles from the changelog section
A few categories stand out in the Full Changelog PR listing (PR-title-level observations; details live in the merged code): FedRAMP API keys route to government endpoints by default (#57178); the GCP project and location are configurable in provider settings (#56560), plus a credential-file override for GCP provider authentication (#57188); the 1 MiB response limit is removed for explicit model catalogs (#56525); Guardian v2 async classification now runs entirely through conversations (#56556), alongside several Guardian truncation and context-composition adjustments; marketplace auto-update failures surface as startup warnings (#57166); SQLite initializes with DELETE journaling on NFS (#56552); masked AWS credentials are supported for local proxy signing (#56554); and MITM hooks gain scoped request-body secret substitution (#57174). The two memory items (#57240, #57568) are covered in our memory v2 explainer.
8. Information boundaries of this batch
What the release notes do not say, and this article will not supply: the actual keymap for transcript navigation (the notes only say configurable), the trigger scope of the clearer safety-block explanations, Terra's regional availability in the Vertex and Bedrock catalogs, or the reason for the PyPI lag. Terra's API-side context and pricing remain marked unverified.
Next steps
- Codex CLI 0.164–0.165, Explained: GPT-6.1 Terra in the Catalog, a Vertex AI Provider, /presets, memory v2 — the previous line, where Terra's bundled-catalog entry and the Vertex AI provider landed
- Codex memory v2 Explained: the memories_v2 Workspace, External-Agent Memory Imports, and memory_summary.md — the mechanism background for this release's two memory PRs
- Codex CLI 0.163.0, Explained: Daybreak On by Default, Instant Interrupts, /config Search — the previous stable-release explainer
- OpenAI Models Release Notes (2026, Living Document) — our continuously updated OpenAI ecosystem timeline
Key points
- Release cadence (GitHub API, checked 2026-10-30): after 0.165.0 (Oct 27, 20:09 UTC), the 0.166.0 line ran six prereleases from alpha.2 (Oct 27, 11:54 UTC) to alpha.6 (Oct 29, 01:41 UTC), then went stable at Oct 29, 19:47 UTC — stable shipped the same day as the last alpha
- GPT-6.1 Terra catalog expansion (#57293): the release notes read Added GPT-6.1 Terra to the bundled, Vertex AI, and Amazon Bedrock model catalogs — our tag-level check shows the bundled models.json on rust-v0.166.0 is byte-identical (same md5) to rust-v0.165.0, so the bundled entry already existed in 0.165 and the new coverage is Vertex AI plus Bedrock. In models.json Terra stays visibility=list, priority 7, context_window 500000 / max_context_window 872000 (a Codex-side runtime window, not an API number)
- Interaction surface: keyboard-driven transcript selection with word, line, and page motions, integrated search, and configurable shortcuts (#57221–#57224); async questions, choices, answers, and transcript replies now render Markdown consistently (#57259/#57292); safety blocks get clearer explanations plus selectable safer follow-up prompts (#57363/#57365)
- Two command surfaces: codex queue --resume resumes unloaded threads after a message is queued (#57232); codex doctor --network-only runs focused configuration, connectivity, and WebSocket diagnostics (#57606)
- Reliability fixes: resume/fork preserves acknowledged approval and permission settings (#57227); reasoning-token accounting survives resume/fork to avoid premature compaction (#57839); local image attachments survive composer-text restoration (#57502); unarchiving a parent restores its descendant threads (#57635); app-server thread startup/shutdown races fixed with prompt runtime release (#57787/#57788); TUI reconnection recovery including offline /quit and /exit (#57820)
- Distribution (checked 2026-10-30): npm @openai/codex and @openai/codex-sdk both list 0.166.0 as latest (posted 2026-10-29T19:56 UTC); PyPI's openai-codex is still at 0.163.0 (posted Oct 15, 21:50 UTC) — 0.164.0, 0.165.0, and 0.166.0 have all skipped PyPI, widening the gap from two versions to three
Frequently asked questions
Official references
- ChangelogCodex CLI 0.166.0 Release Notes (GitHub, 2026-10-29T19:47 UTC, checked via the GitHub API)
- Docsmodels.json (openai/codex rust-v0.166.0 tag, fetched 2026-10-30, byte-identical to rust-v0.165.0)
- Changelog@openai/codex (npm registry, latest 0.166.0, checked 2026-10-30)
- Changelog@openai/codex-sdk (npm registry, latest 0.166.0, checked 2026-10-30)
- Changelogopenai-codex (PyPI, latest still 0.163.0, checked 2026-10-30)
Related articles
Codex CLI 0.164–0.165, Explained: GPT-6.1 Terra in the Catalog, a Vertex AI Provider, /presets, memory v2
Two stable releases since 0.163: 0.164.0 (Oct 23) adds /presets and inline /model and /permissions; 0.165.0 (Oct 27) ships a Vertex AI provider and GPT-6.1 Terra in the catalog. Checked against the official notes.
Read articleOpenAI Maintains Software with GPT-6: 1,400+ PRs, 800k Lines, 92%-for-14%
OpenAI's Core Experimentation team details GPT-6-powered maintenance: 1,400+ dead-code PRs, stale-experiment retirement, and a comparison where 6.1 Sol Ultra matched 92% of Astra's findings at 14% of the cost.
Read articleCodex CLI 0.163.0, Explained: Daybreak On by Default, Instant Interrupts, /config Search
The 0.163 preview line closes into a stable release: Daybreak on by default for eligible accounts, instant interrupts by default, /config preference search, a leader key, parallel read-only tools — and the bundled zsh is gone.
Read articleSubscribe to GPTMap Weekly
One email every Monday: curated OpenAI updates, deep dives, and best practices. No ads, unsubscribe anytime.
Submitting opens Buttondown in a new tab to confirm your subscription.