What Changed in Codex CLI 0.153.0 Through 0.153.2

Three releases, one day apart

Codex CLI shipped 0.153.0, 0.153.1 and 0.153.2 in quick succession, all dated 3 September 2026 per OpenAI's own release notes on the openai/codex GitHub repository, the primary source this article draws from directly. developers.openai.com, which would carry fuller configuration reference for some of what's below, was not reachable while researching this piece, so anything not confirmed in the release notes themselves is flagged rather than guessed at.

The two patch releases, 0.153.1 and 0.153.2, exist almost entirely to backport and then correct support for a new OpenAI model. The substantive terminal and workflow changes are in 0.153.0 itself.

0.153.0: editor and workflow changes

Vim mode gets real undo and redo. The release note: "Vim mode now supports undo with u and redo with Ctrl+R, preserving complete drafts including pasted content and attachments." That last clause matters more than it looks. A draft in a terminal agent often isn't just typed text, it can include pasted code blocks, file attachments, or multi-line content assembled over several edits. Undo and redo that only tracked plain keystrokes would be far less useful than one that preserves the whole draft state, attachments included, across each step.

The plugin CLI can manage remote marketplaces directly. "The plugin CLI can list, install, and remove plugins from remote marketplaces" is the full release note. Previously, working with a plugin marketplace meant more manual setup; this puts list, install and remove behind ordinary CLI subcommands against a marketplace hosted somewhere other than your local machine. If you're already using Codex CLI's /import command to bring skills across from another agent, this is the adjacent piece: once a skill or plugin exists in a marketplace, you no longer need bespoke steps to pull it into a fresh Codex CLI install.

tui.auto_recap = false gives you an off switch for automatic recaps. Codex CLI has offered a /recap command for a while; this setting lets you turn off the automatic version specifically, "while keeping manual [recap] functionality." If you found automatic recaps interrupting a long session with a summary you didn't ask for, this is the fix, without losing the ability to trigger one yourself when you actually want it.

TUI history stopped summarising as much. The release note: "TUI history shows complete patches, input sent to background terminals, and individual completed commands." Before this, history views in the terminal UI evidently condensed some of that detail. Now a patch shows in full rather than as a truncated diff, and background terminal input and individual completed commands get their own visible entries instead of being folded into a summary line. This is a straightforward transparency improvement for anyone reviewing what an agent actually did across a session, particularly one running multiple background commands at once.

Plus and Team users get an earlier low-allowance warning. "Plus and Team users receive an earlier warning when less than half of their allowance remains," per the release note. No further detail on the exact threshold mechanics was given beyond that.

0.153.0's bug fixes worth knowing

A handful of fixes in this release affect how Codex CLI behaves around review and reconnection, rather than adding new surface area:

• Guardian reviews are skipped for confirmation-only actions in Full Access mode. Codex CLI's Guardian layer evaluates tool-call requests automatically, escalating genuinely risky ones rather than asking for approval on everything. The pull request implementing this change states the reasoning plainly: "Full Access already combines approvalPolicy: 'never' with unrestricted permissions, so confirmation-only actions do not require a model review." In other words, if Full Access mode has already decided nothing needs asking, running a Guardian review on top of that for a merely confirmation-shaped action was redundant work. The same PR preserves existing handling for cancellations, explicit denials, and anything requiring actual user input, none of that changed.
• Guardian review history survives compaction, restarts, and user-created forks. Previously, some combination of those events could apparently lose that history; it's now retained across all three.
• Remembered MCP tool approvals are scoped to the selected app account. An approval you granted under one connected account no longer applies when you've switched to a different one.
• Rollout compression includes shared histories, and codex exec resume now handles compressed rollouts correctly.
• TUI sessions reconnect after an external app-server connection drops, preserving drafts and transcripts rather than losing session state.

Codex CLI's own documentation for exactly how Guardian's scoring model works, what counts as "confirmation-only" beyond the examples above, wasn't available while researching this piece. Treat the mechanism as real and specifically changed in this release, without assuming more about its internals than the release notes and linked PR actually state.

Guardian versus Claude Code's auto mode classifier

Codex CLI's Guardian and Claude Code's auto mode classifier solve a version of the same problem from different angles, which is worth naming even though the two tools describe their own mechanisms differently. Both exist to reduce how often a routine, low-risk tool call needs a human to click approve, and both escalate the calls that look genuinely risky rather than treating every action the same way. Where they diverge, based on what's publicly documented for each, is scope: Claude Code's classifier is described in terms of specific named rules, a built-in data-exfiltration rule, a Containment Escape rule added in v2.1.257, evaluated against your project's own auto-mode configuration. Codex CLI's release notes describe Guardian in terms of categories of action, "confirmation-only" versus one requiring genuine review, without publishing the equivalent named-rule structure. Neither company's own reference documentation for the internals of either system, developers.openai.com for Guardian, was reachable while researching this piece, so this comparison is necessarily at the level of what each project's own release notes and changelogs say about behaviour, not a claim that the two mechanisms work identically underneath.

0.153.1 and 0.153.2: catalog support for a new model

0.153.1's entire release note: "Added support for configuring GPT-6-Astra through the API without changing the default model or showing it in the model picker." The linked changelog entry names it directly: "Backport GPT-6-Astra model catalog to 0.153."

That's a narrow, specific claim, and it's worth being precise about what it does and doesn't confirm. It says Codex CLI can now be configured to use a model called GPT-6-Astra via its API-level configuration. It doesn't change what you see by default: no automatic switch, and the model doesn't show up in the interactive model picker unless you've configured it. It says nothing about GPT-6-Astra's pricing, context window, benchmark results, or availability tiers, those are facts about the model itself rather than about Codex CLI, and they belong on OpenAI's own announcement page. That page was not reachable while researching this piece, so this article doesn't repeat figures reported elsewhere that weren't opened and confirmed directly.

0.153.2, released the same day, fixed a display-only issue: "Corrected the GPT-6-Astra Fast tier description to say '2x speed, increased usage' instead of '1.5x.'" The release notes are explicit that this "affects only displayed text, not actual request processing", so if you configured the Fast tier under 0.153.1, nothing about how your requests were actually served changed, just the number shown to describe it.

Together, these two patches confirm two things Codex CLI's own release notes state directly: GPT-6-Astra exists as a configurable model option starting with this line, and it has at least a standard and a "Fast" tier, the latter running at roughly twice the speed at increased usage cost, per the corrected description. Anything beyond that about the model itself needs a source this routine could actually open, which OpenAI's own page wasn't.

Every change in one table

Reading straight down that list, the pattern is a release focused on polish and trust in existing systems, undo that actually preserves what you pasted, history that shows the real patch rather than a summary, review logic that stops asking twice when a mode has already decided nothing needs asking, rather than a release built around one headline feature. The model catalog addition in the two patches is the exception, and even that shipped as a narrowly-scoped backport rather than a 0.153.0 launch feature.

What this means if you're on an older version

If you're running 0.152.0 or earlier, none of the Vim undo/redo, plugin marketplace management, tui.auto_recap, or GPT-6-Astra catalog support exists yet in your install. codex --version tells you what you're on; comparing against OpenAI's release notes for each version between yours and 0.153.2 tells you exactly what you'd gain by upgrading. If you're specifically interested in the model catalog change, 0.153.1 is the minimum version, not 0.153.0.

This release line sits right after the hooks changes covered across 0.148.0 through 0.152.0 and the planning tool's switch to opt-in in 0.152.0. Neither of those carried forward into new behaviour in 0.153.x; this line's changes are additive on top of them, not a continuation of the hooks work specifically.

Troubleshooting

Vim undo isn't restoring a pasted attachment. Confirm you're on 0.153.0 or later; earlier versions don't have draft-preserving undo at all, only plain text undo where it existed.

GPT-6-Astra doesn't appear in the model picker after upgrading. That's expected behaviour per the 0.153.1 release note: it's configurable through the API but deliberately not shown in the picker or set as default. Configure it explicitly rather than expecting it to appear automatically.

A Guardian review still runs for an action you expected Full Access to skip. The skip applies specifically to confirmation-only actions. Cancellations, explicit denials, and anything requiring actual user input keep their existing review behaviour regardless of Full Access, per the linked pull request.

An MCP tool approval you set earlier isn't being honoured. Check which app account is currently selected. As of 0.153.0, remembered approvals are scoped per account, so switching accounts means re-approving tools you'd already cleared under a different one.

Where to go next

For the releases immediately before this line, see Codex CLI's interactive task dashboard and what changed in 0.149.0 and Codex CLI hooks: the full timeline from 0.148.0 to 0.152.0. For getting skills into Codex CLI in the first place, see how to install skills in Codex CLI and Claude Code vs Codex CLI for skills. Browse the current skill catalogue at getclaudeskills.com/skills.