OpenAI Codex release notes: Guardian fixed and iOS tasks made clearer

OpenAI Codex received two fixes that affect the review chain itself. ChatGPT for iOS 1.2026.188 makes Codex tasks easier to follow and ensures the first prompt respects the selected approval preset. Codex CLI 0.144.2 also restores the Guardian flow after a prompting regression. Together, they touch two common handoffs: mobile to task, and code change to review.

OpenAI Codex release notes: two review flows are fixed

Codex CLI is OpenAI's local coding agent for reading, changing, and testing projects. Guardian auto-review is Codex's automated review flow, built around a policy, a review-request format, and dedicated tools.

Version 0.144.2 restores the previous Guardian policy, request layout, and tool behavior. OpenAI describes it as a full rollback of a prompting update on the 0.144 release branch, with no other product changes. This is a correction to review behavior, not a new review feature.

Source: OpenAI Codex changelog, CLI 0.144.2

Source: GitHub PR #32672, Guardian prompt rollback

Codex tasks are clearer in ChatGPT for iOS

ChatGPT for iOS 1.2026.188 adds inline visualizations to Codex tasks. The release also improves links to newly created tasks, tool activity, progress indicators, and feedback when opening files.

Two fixes matter for handoff. Fast mode should now be selected and restored per task, and the first prompt should no longer ignore the selected approval preset. OpenAI does not document a new menu path or gesture for these changes, so we are not adding unverified UI instructions.

Source: OpenAI Codex changelog, ChatGPT for iOS 1.2026.188

0.144.4 is the current stable Codex CLI

After the Guardian fix, OpenAI published 0.144.3 as a version-only release and 0.144.4 with no user-facing changes. The current stable release is therefore 0.144.4, while the practical behavior change in this coverage window comes from 0.144.2.

Source: OpenAI Codex changelog, CLI 0.144.4

Human step: update the CLI before testing

OpenAI's current release notes show this installation command:

npm install -g @openai/[email protected]

Run the update as a human action. The agent prompt below reviews a change; it should not install packages or alter approval settings.

Source: OpenAI Codex changelog, installation line for 0.144.4

Short example: use the new Codex feature

Review the current diff without editing files. Return findings in severity order with file and line evidence. State which checks you ran, what remains uncertain, and whether the change is ready for human approval.

Good output should:

• Provide concrete findings with file and line references
• Separate verified defects from uncertainty
• Report completed checks without implying that other tests ran
• Leave the change untouched until a person decides what happens next

What a Nordic team can test now

Choose a pull request that already has human review and run the same read-only review prompt in the updated CLI. If you use iOS, also confirm that the task link opens the right Codex task and that the selected approval preset applies from the first prompt. Keep findings, completed checks, and the decision in one review receipt.

You can keep this inside the review process you already use. In Tool Forge, the same receipt can connect the pull request, approval point, and audit log so the review remains traceable even when work starts on mobile.