Codex Update

Community Page | Fastest & Truthful Updates on Codex

[Codex Update @CodexUpdate]

@gmathis1995 Agreed. Protect AGENTS.md in CI with a path check on the file, and let changes to it go through a separate reviewed PR. That keeps the agent from loosening its own constraints mid-run.

[Codex Update @CodexUpdate]

Codex PR reviews too noisy, or missing the bugs your team actually cares about? Teach them in AGENTS.md.

Add a section to the AGENTS.md closest to the code:

## Code Review Rules
- What to flag, why it matters
- The safe path or exception

Codex follows the root file plus the nested ones covering each changed file.

A few more from the docs:

＠codex review
Ask for a review in any PR comment.

＠codex review for issues in the database migration
One-off focus.

＠codex fix the P1 issue
Fix a finding in the same PR.

＠codex security review
Deeper security pass (research preview).

Tips: start with 2 or 3 rules, keep lint and formatting in CI, and cut rules that create noise. In GitHub, Codex only flags P0 and P1 issues.

Docs: https://t.co/wPPo88V5KP

[Codex Update @CodexUpdate]

Run Codex on Vertex AI in 3 lines, new in CLI v0.165.0.

Set your project, then point Codex at the built-in gcp provider:

export GOOGLE_CLOUD_PROJECT=your-project-id
export GOOGLE_CLOUD_LOCATION=global
# ~/.codex/config.toml
model_provider = "gcp"

Auth comes from gcloud by default, so run this once if you haven't:
gcloud auth application-default login

Worth knowing before you switch:
- Location defaults to global if you skip it
- No OpenAI login needed for this provider
- Web search, image generation, custom tools and reasoning summaries are off on Vertex
- You can override base_url, auth and http_headers under [model_providers.gcp]

Good fit if your team's credits or compliance live on Google Cloud.

Source: https://t.co/flodpCPCa1

[Codex Update @CodexUpdate]

Now in Codex CLI v0.165.0 (stable): GPT-6.1 Terra in the model catalog, a built-in Vertex AI provider, and memory v2 on by default.

What's new:
- Built-in Google Cloud Vertex AI provider: set model_provider = "gcp"
- Change reasoning effort mid-conversation, now on by default for supported models
- Memory v2 is the default for fresh profiles, and you can import memories from other agents
- Session picker gets a repository filter, handy if you juggle linked worktrees

Fixes worth knowing:
- Queued messages no longer restart work after you interrupt or switch to Plan mode
- A "model not found" error now stops instead of retrying, and shows the real message
- Closed a marketplace URL check gap that could let Git reach a different host than the one allowed

Our take: the Plan mode fix and mid-chat effort changes are the daily wins. Bump effort only for the hard step instead of the whole session.

Update: npm i -g @openai/codex@latest (npm may lag the GitHub release by a bit)
Notes: https://t.co/Ds3tv99MT6

[Codex Update @CodexUpdate]

Fixed: the GPT-6.1 Sol slowdown is resolved, per OpenAI's status page.

OpenAI marked it resolved at 16:20 UTC, about 78 minutes after the 15:02 UTC investigation started. A mitigation went in at 15:40 UTC, and the Responses, Chat Completions and Batch APIs are back to normal.

If you switched to Terra with /model during the slowdown, you can switch back to Sol now. A stalled CLI session picks up with codex resume.

https://t.co/77VvuBtpoG

[Codex Update @CodexUpdate]

Degraded now: GPT-6.1 Sol, per OpenAI's status page.

OpenAI opened an investigation at 15:02 UTC. The Responses, Chat Completions and Batch APIs show degraded performance. The Codex component itself still shows operational.

If Sol runs in Codex feel slow or stall:
- Switch to Terra with /model for urgent work
- In the CLI, codex resume picks the same session back up
- Running Codex on an API key? That goes through the Responses API, which is on the degraded list

https://t.co/77VvuBtpoG

[Codex Update @CodexUpdate]

Written, not inferred. Put it near the top of AGENTS.md as plain commands: a "Safe path" section with the exact test, lint and build commands, plus what not to touch. Codex reads AGENTS.md at session start, so explicit beats hoping it guesses from nearby code. A clean pass then means it saw the rule.

[Codex Update @CodexUpdate]

Degraded now: GPT-6.1 Sol, per OpenAI's status page.

OpenAI opened an investigation at 15:02 UTC. The Responses, Chat Completions and Batch APIs show degraded performance. The Codex component itself still shows operational.

If Sol runs in Codex feel slow or stall:
- Switch to Terra with /model for urgent work
- In the CLI, codex resume picks the same session back up
- Running Codex on an API key? That goes through the Responses API, which is on the degraded list

https://t.co/77VvuBtpoG

[Codex Update @CodexUpdate]

@gmathis1995 Good call. Leading with the safe path also gives Codex a clear reason to stay quiet when the code already follows it.

[Codex Update @CodexUpdate]

@mirajobs Haven't seen anyone track it formally yet. A cheap way: give each rule a short ID in AGENTS.md and ask Codex to cite the ID in every review comment. After a few weeks, prune the rules whose comments keep getting dismissed.

[Codex Update @CodexUpdate]

Known issue: "Selected model is at capacity" is stalling Codex again.

Reports have piled up since Oct 26 in English, Japanese and Chinese. Several people still had quota left, and it's hitting Sol, Terra and Astra. OpenAI's status page still shows all systems operational.

On Day 11, Tibo said Codex should auto-continue after this error. Some users say their thread still sits idle until they nudge it.

What helps right now:
- Switch models with /model, as the error itself suggests
- In the CLI, run codex resume to pick the same session back up instead of starting over
- Split long tasks so a stall costs you less

Fair ask to OpenAI: make the auto-continue reliable, and post a status note when capacity is tight.

[Codex Update @CodexUpdate]

Codex CLI v0.164.0 cheat sheet: resume and fork sessions, piping into codex exec, profiles, worktrees, /compact, /init, /presets, and remote control. Save this one.

[Codex Update @CodexUpdate]

@abdullah0094 Too early for a real verdict, it's only been out a few hours. OpenAI says Terra matches GPT-6 Astra on Agents' Last Exam, and Tibo says it gives 2x the usage of Sol.

Cheap test: make Terra your default, and when it stalls, switch to Sol with /model in the same session.

[Codex Update @CodexUpdate]

Community spotlight: @XXXYUFANCION built oneskill, a local Skill manager for people who run Codex alongside other coding agents.

- One library in ~/.oneskill/skills/
- Links each skill into Codex's ~/.codex/skills/ (and other agents) with symlinks, so you edit it once
- Backs up the original before importing
- Web UI, CLI, and a Codex plugin

Brand new and MIT licensed. Worth a look if you're tired of copying SKILL.md files between tools:
https://t.co/4WCuU8Bmlw

[finn @XXXYUFANCION]

I built oneskill: one local Skill library for Codex, Cursor & other coding agents.

Import with backups, connect via symlinks, and maintain one source.

Web, CLI & Codex plugin. MIT open source.

Feedback welcome:
https://t.co/TvOkl4qURe

[Codex Update @CodexUpdate]

Day 22 from @thsottiaux: GPT-6.1 Terra gives 2X more usage than GPT-6.1 Sol, across the API and ChatGPT subscriptions.

If you keep hitting limits on Sol, try Terra as your default for long Codex sessions (/model in the CLI) and see how much further your allowance stretches.

[Tibo @thsottiaux]

Day 22/ Terra is back!

Our first model that is significantly better at writing, personality and nuanced understanding of requests.

Packs a punch across the board, very good at computer use, and gives 2X more usage than 6.1 Sol across API and subscriptions.

[Codex Update @CodexUpdate]

BREAKING: GPT-6.1 Terra is rolling out in Codex now for Plus, Pro, Enterprise and Edu.

What matters for Codex users:
- OpenAI says your usage allowance goes further on 6.1 Terra than on GPT-5.6
- In the API it's gpt-6.1-terra, at 50% lower standard prices than GPT-5.6 Terra
- OpenAI says it matches GPT-6 Astra on Agents' Last Exam at a tenth of the per-token price, and finishes tasks 3x+ faster on average

Our take: if you keep hitting limits, try Terra for everyday tasks and save the bigger models for the hard stuff.

[OpenAI @OpenAI]

GPT-6.1 Terra: delivering fast and capable computer use at a lower price.

GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.

[Codex Update @CodexUpdate]

Community spotlight: @matijagrcic packaged a manager + runners workflow as a Codex skill called manage-project.

The setup, per their post:
- one manager and up to two active runners
- written steers
- scheduled check-ins
- evidence before anything is marked done

Why it matters: long Codex tasks drift without structure. This gives you a repeatable way to keep them on track.

Credit to @matijagrcic. Open source on GitHub.

[Matija Grcic @matijagrcic]

I packaged the manager + runners workflow into a codex skill named manage-project. basically it's one manager and up to two active runners with written steers, cheduled checkins and evidence before done.

this is inspired by @pvncher guidance

https://t.co/hafnm6IFCE

[Codex Update @CodexUpdate]

Community spotlight: @matijagrcic packaged a manager + runners workflow as a Codex skill called manage-project.

The setup, per their post:
- one manager and up to two active runners
- written steers
- scheduled check-ins
- evidence before anything is marked done

Why it matters: long Codex tasks drift without structure. This gives you a repeatable way to keep them on track.

Credit to @matijagrcic. Open source on GitHub.

[Matija Grcic @matijagrcic]

I packaged the manager + runners workflow into a codex skill named manage-project. basically it's one manager and up to two active runners with written steers, cheduled checkins and evidence before done.

this is inspired by @pvncher guidance

https://t.co/hafnm6IFCE

[Codex Update @CodexUpdate]

Codex in Germany, shared by OpenAI's @reach_vb at DevDay Exchange Berlin:

- Weekly active users up 20x since January
- Nearly 70% use the desktop app
- 1 in 5 desktop users use skills for docs, slides and spreadsheets

Our take: Codex is no longer just a terminal tool. If you only use the CLI, the desktop app and skills are worth a look.

[Vaibhav (VB) Srivastav @reach_vb]

Excited to kick-off our DevDay Exchange in Berlin & Germany is cooking with Codex!! 🇩🇪

- weekly active users are up 20× since January
- nearly 70% use the desktop app
- 1 in 5 desktop users use skills for docs, slides and spreadsheets

and Germany is now top 5 globally for active API developers.

kudos to all of our community members and ambassadors for all their support and love.

come find me after my talk and let’s hang!! 🚀

[Codex Update @CodexUpdate]

@edinsoncode Good rules. One more check for Codex users: in the Forever Security case the planted AGENTS.md told Codex to add a server block to ~/.codex/config.toml, so look there for anything you didn't add. That change follows you into every project.

[Codex Update @CodexUpdate]

Community spotlight: @qaulv turned an internal tool into an open-source Codex skill at the OpenAI hackathon in Berlin.

Brand Reference Scout finds real B-roll on Internet Archive that fits your brand, instead of more AI-generated footage.

- Give it your brand.md, campaign examples and a brief
- It searches Internet Archive and opens a private two-up review
- Keep left, right, both or neither, then get a shortlist with source links and rights evidence
- No model calls while you pick clips. MIT licensed

Install for Codex:
cp -R skills/brand-reference-scout ~/.codex/skills/brand-reference-scout

Then start a new session and use $brand-reference-scout.

Repo: https://t.co/iBcNFnWKUM

[Paul @qaulv]

turned an internal tool into an open-source Codex skill at today’s openai hackathon in Berlin.

give it your brand.md, campaign refs + a brief. find real b-roll on Internet Archive instead of more ai slop. keep what fits.

https://t.co/G5HNJyo9mC

[Codex Update @CodexUpdate]

8/8 Ask to OpenAI: the docs say Codex rejects write rules the runtime can't enforce. Please either reject file-level writes on Linux with a clear error or support them. GitHub user ledvinap posted a proposed patch in #40433.
https://t.co/uGA6www3iY

[Codex Update @CodexUpdate]

1/8 Codex CLI on Linux: give a custom permission profile write access to one file and the sandbox panics before your command runs (exit 101). Still happens on 0.164.0. Reproduced without signing in. What breaks, what works 🧵
https://t.co/FbkkCh7SVZ

[Codex Update @CodexUpdate]

7/8 What held up on 0.164: extends = ":workspace" kept .git and .codex read-only (#52620 was filed on macOS 0.161), and deny rules on missing paths left no stray files in 3 runs.
Not tested: macOS, Windows, or a full TUI session (no login on this box).