Quick answer
OpenAI released Codex CLI 0.153.4 on September 4 as the latest hotfix in the 0.153 series. It fixes GPT-6 Astra's visibility in the bundled model picker and makes Astra the bundled default when a user has not explicitly configured another model. It also limits Astra's asynchronous-question guidance to sessions where that tool is available. The preceding 0.153.3 hotfix added Astra to the Amazon Bedrock picker for Mantle and Runtime global and U.S. routes. The underlying 0.153 release added remote plugin marketplace management, resilient TUI reconnection, account-scoped MCP approvals, richer terminal history, and an off-by-default experimental context-management mode. These Codex defaults do not verify Astra availability in Chat AI.
Codex CLI 0.153.4 shows Astra and uses it as the bundled default
The 0.153.4 hotfix repairs GPT-6 Astra's visibility in the bundled model picker and makes it the bundled default when no model is explicitly configured. An existing explicit model choice should therefore remain the controlling setting, while an unpinned session can inherit the new default. The release notes do not say this overrides account eligibility or model access, so seeing a default in the client should not be treated as proof that every account can run it.
Sources: OpenAI on GitHub
Astra now gets asynchronous-question guidance only when the tool exists
OpenAI also changed Astra's guidance so it recommends asynchronous questions only in sessions where the relevant tool is available. This follows the 0.153.3 correction that pointed Astra to the supported text-only clarification tool. Both changes concern the client guidance supplied around the model; the release notes do not claim a change to Astra's model weights or general reasoning capability.
Sources: OpenAI on GitHub, OpenAI on GitHub
Codex CLI 0.153.3 added Astra to supported Amazon Bedrock routes
The preceding 0.153.3 hotfix added GPT-6 Astra to the Amazon Bedrock model picker for Mantle and Runtime global and U.S. routes. Eligible Codex users working through those routes can select the model from the catalog, but the release notes do not announce changes to AWS permissions, regional eligibility, model entitlements, or rollout access.
Sources: OpenAI on GitHub
Codex CLI 0.153 expands plugin and terminal workflows
The stable 0.153.0 release lets the plugin CLI list, install, and remove plugins from remote marketplaces. The terminal interface also gains Vim-style undo with u and redo with Ctrl+R, preserves pasted content and attachments in draft history, and shows complete patches, background-terminal input, and individual completed commands in its history. These are client and workflow changes rather than a model upgrade.
Sources: OpenAI on GitHub
Interrupted TUI sessions can reconnect without silently sending queued work
OpenAI says Codex now reconnects a terminal session when its external app-server connection drops. Drafts and transcripts are preserved, while submissions whose state is uncertain or that remain queued stay paused for review. That distinction matters after a network or process interruption: users can inspect the recovered session before deciding whether to resend or continue.
Sources: OpenAI on GitHub
Remembered MCP approvals are now tied to the selected app account
The release scopes remembered MCP tool approvals to the app account selected for that connection. This reduces the chance that an approval remembered for one linked account is treated as applicable to another account. OpenAI also fixed startup of MCP servers that use relative executable paths on macOS. Teams with multiple accounts or locally configured servers should verify both account selection and server startup after upgrading.
Sources: OpenAI on GitHub
Guardian review behavior now follows the active approval mode more closely
In Full Access mode, confirmation-only actions no longer run a separate Guardian review. In User approval mode, background Guardian scoring and prewarming are skipped. The release notes say sensitive-action checks and requests for user input keep their existing handling, so these changes should not be read as removing those boundaries. Guardian review history also survives compaction, restarts, and user-created forks while respecting rollback and subagent isolation boundaries.
Sources: OpenAI on GitHub
Recaps, usage warnings, and history get practical controls
Users can set tui.auto_recap to false to stop automatic recaps while retaining the manual /recap command. Plus and Team users also receive an earlier warning when less than half of their allowance remains in an approximately five-hour usage window. The warning describes remaining allowance; it does not change the plan limit. Teams that depend on automatic summaries should check their configuration before turning recaps off.
Sources: OpenAI on GitHub
Experimental context management remains disabled by default
Version 0.153 adds features.context_management.experimental_mode. When explicitly enabled for eligible ChatGPT Plus, Pro, or Pro Lite sessions using the Codex backend, it activates token-budget context, history notes, and the new_context tool. API-key sessions, custom providers, and temporary structured threads are excluded. Because the setting is experimental and off by default, users should not assume those behaviors are active after a routine upgrade.
Sources: OpenAI on GitHub
Upgrade with a short plugin, reconnect, and permission check
After installing the release, list configured plugin marketplaces and confirm that a test plugin can be installed and removed from the intended source. Then interrupt a non-critical TUI session to confirm its draft and transcript return without an uncertain request being sent automatically. If you use multiple linked MCP accounts, verify that approvals stay with the selected account. Finally, confirm recap and usage-warning behavior matches your configuration. These are practical checks derived from the release notes, not an OpenAI certification procedure.
Sources: OpenAI on GitHub
A Codex default does not establish Chat AI availability
Codex CLI 0.153.4 updates OpenAI's coding-agent client and its bundled default. That does not add Astra to Chat AI or verify that the model is available there. Chat AI availability is therefore marked not applicable for this tooling article.
Sources: OpenAI on GitHub
Frequently asked questions
What readers usually ask
What changed in Codex CLI 0.153.4?
The hotfix makes GPT-6 Astra visible in the bundled model picker, uses it as the bundled default when no model is explicitly configured, and limits asynchronous-question guidance to sessions where the tool is available.
Will Codex CLI 0.153.4 replace an explicitly configured model?
The release notes say Astra becomes the bundled default when no model is explicitly configured. They do not say an explicit model selection is overridden.
Can Codex CLI 0.153 install plugins from remote marketplaces?
Yes. OpenAI's release notes say the plugin CLI can list, install, and remove plugins from remote marketplaces.
What happens if a Codex TUI app-server connection drops?
The session can reconnect with drafts and transcripts preserved. Submissions that are uncertain or still queued remain paused for review.
How do I turn off automatic recaps in Codex CLI 0.153?
Set tui.auto_recap to false. The manual /recap command remains available.
Is the new context-management mode enabled automatically?
No. features.context_management.experimental_mode is disabled by default and has documented account, backend, provider, and thread-type eligibility limits.
Does Codex CLI 0.153.4 make GPT-6 Astra available in Chat AI?
No. The release changes Codex's bundled model picker and default. It does not verify Astra availability in Chat AI.
Evidence