OpenAI Codex 0.153.0 strengthens session recovery, plugins, Guardian history and context controls

OpenAI’s Codex 0.153.0 release focuses on the operational details that determine whether an AI coding session remains understandable and recoverable over time. The update expands terminal editing and history, brings remote marketplace operations to the plugin CLI, improves recovery from app-server disconnections and makes Guardian review evidence more durable. It also refines MCP approval scope, compressed rollout handling and app-server metadata.

The supplied GitHub snapshot identifies the repository as public and the release as 0.153.0, with the tag rust-v0.153.0 and revision 41e22fe. It says github-actions published the release at 03 Sep 01:37 and reports 83 commits to main since the release. The same snapshot displays approximately 18.6k forks and 121k stars, as well as 162 assets and 14 reactions. These are captured GitHub page values, not measurements of the release’s quality, performance or adoption.

New Features

Vim undo and redo preserve the complete composer state

Vim mode in the terminal composer now supports undo with u and redo with Ctrl+R. The release makes the scope of that history important: the controls preserve complete drafts, including pasted content and attachments. The undo and redo changes are associated with pull requests #41941 and #42140.

This is accompanied by a more complete TUI history. Codex can show full patches, input sent to background terminals and individual completed commands. The changelog separates that work into #41893, which displays successful commands individually, and #42107, which exposes full patches and terminal input. The result is a session record containing more of what the user and tools actually did.

Users can also change recap behavior without losing access to recaps altogether. Setting tui.auto_recap = false turns off automatic recaps, while the manual /recap command remains available. That configuration addition is tracked in #42101.

Remote marketplaces become manageable through the plugin CLI

The plugin CLI can now list, install and remove plugins from remote marketplaces through the work recorded in #42150. Related changelog entries cover a plugin-reconciliation app-server API in #41949, marketplace source policy for curated plugins in #41953, centralized remote mutations in PluginsManager in #42114, Git marketplace upgrades in #42149 and a preference for remote Sites over the bundled plugin in #42100.

The source does not provide example commands, identify every compatible marketplace or describe a migration procedure. The verified change is therefore the availability of list, install and remove operations for remote marketplaces, supported by the related internal marketplace work in the changelog.

Allowance warnings arrive sooner for two plans

ChatGPT Plus and Team users now receive an earlier allowance warning. It is shown when less than half of the user’s allowance remains in an approximately five-hour usage window. The change is associated with #42142. The release supplies neither a fixed request count nor a token threshold, so the numerical details that can be stated are the below-half trigger and the approximately five-hour window.

Bug Fixes

TUI sessions reconnect without silently resubmitting uncertain work

Codex TUI sessions can reconnect after an external app-server connection drops. Drafts and transcripts are preserved, and navigation is restored after reconnection. The recovery behavior deliberately leaves submissions with an uncertain state, along with queued submissions, paused for user review. The relevant work spans #41911, #41916 and #41918.

The broader changelog contains several complementary continuity changes. MCP event subscriptions remain alive after task unloading, an event-stream manager was added, realtime conversation history can be recorded in Core and response token usage persists in rollout history. TUI status timing survives when the status row is hidden, transcript layout caches survive backtrack selection and resume guidance in exit summaries was clarified. Recent sessions can also be shown in the agent command center.

Guardian processing changes according to approval mode

Version 0.153.0 changes when Guardian review work runs. In Full Access, confirmation-only actions skip Guardian reviews under #42147. In User approval mode, background Guardian scoring and prewarming are skipped under #42256. The release explicitly preserves the existing treatment of sensitive-action checks and requests for user input, so the change should not be described as removing those controls.

Guardian history is also made more durable. Review evidence survives compaction, restarts and forks created by users, while continuing to respect rollback boundaries and isolate subagent history. The highlighted changes are tied to #41879 and #42065. Related entries cover shared transcript collection, user-message retention, improved diagnostics, failed-review evidence in diagnostic reports and centralized Guardian context composition. Reviews can also be attributed to nested REPL calls and OpenAI app tools, while #42144 adds Guardian V2 analytics events.

MCP approvals are tied to the selected account

Remembered MCP tool approvals are now scoped to the selected app account through #42133. This means the remembered approval is associated with the relevant account link instead of being treated as unqualified session-wide permission. The changelog also includes per-account app approval settings, explicit account selectors for Apps tool calls, app-link settings for MCP approvals and app-link metadata in approval requests. Separate work prepares MCP connections for coordinated OAuth refresh.

On macOS, relative executable paths for MCP servers start more reliably through #42117. The source characterizes this as a reliability fix; it does not say that every possible MCP startup failure has been eliminated.

Compressed histories work across more resume and fork paths

Rollout compression now includes shared histories under #42039. In addition, codex exec resume can handle compressed rollouts when choosing a session by working directory, and #42135 enables thread forks from symlinked session roots. These are distinct compatibility paths involving shared history, working-directory selection and symlinked storage locations.

Other changelog entries avoid scanning archived rollouts while archiving threads and tag Codex home-size metrics with compression state. The release offers no compression ratio, storage benchmark, recovery-time measurement or performance comparison, so these changes are best understood as coverage and reliability improvements rather than quantified efficiency gains.

Configuration and API Updates

App-server metadata exposes optional model settings

App-server thread metadata now includes nullable model and reasoningEffort fields through #42151. Their nullable status matters to API consumers: neither field should be assumed to contain a value in every thread record.

Structured asynchronous questions are supported through request_user_input_async under #42178, but only when the model catalog enables the capability. The release does not say that every model or session can use it.

Additional infrastructure work includes per-call sideband endpoints for existing realtime calls, a reusable OTEL trace WebSocket crate, a cacheable Bazel app-server schema bundle and support for header injection in network requirements. Tool JSON Schema code was divided into more focused modules, while app-tool analytics can record result sources.

The paste-burst setting moves under the TUI namespace

tui.disable_paste_burst replaces the earlier top-level configuration key as part of #41976. The old top-level form remains supported as a fallback. This is therefore a configuration relocation with backward-compatible fallback behavior in this release, not an immediate removal of the prior form.

Experimental context management is off by default

The new features.context_management.experimental_mode setting, introduced through #42385, is disabled by default. When it is enabled for an eligible session, it activates token-budget context, history notes and the new_context tool.

Eligibility is restricted to ChatGPT Plus, Pro or Pro Lite sessions using the Codex backend. API-key sessions, sessions using custom providers and temporary structured threads are excluded. The release does not provide comparative results for the experimental behavior or promise wider availability, making the default-off state, backend requirement and exclusions essential caveats.

Changelog

The full changelog extends beyond the highlighted features and fixes. Native-voice work includes pinned source preparation, a dependency build recipe, helper and installed-host lifecycle foundations, a Windows dependency-build fix and Windows MXC availability recording. Code Mode also omits undersized WAV output.

Developer maintenance includes repository-wide Rust formatter discovery, Python formatting and moving the configuration-schema generator into a dedicated crate. Bundled cleanup hooks are treated as built-ins, while installation detection covers standalone macOS CLI bundles and Vite+-managed Codex installs. Git-root discovery for metadata enrichment is bounded, descriptive labels on local-file links are preserved and plugin-catalog test helpers ignore non-JSON files.

Permission and execution changes make transforms and preapproval aware of executor-path context, resolve permission requests in that context, apply remote-platform semantics to execution safety checks and report the configured sandbox policy consistently. The changelog also expands extension-permission regression coverage and allows the approval reviewer for an active turn to be updated.

Observability work improves tracing for nested tool calls and execution processes, emits turn-cost telemetry for ChatGPT sessions, retains raw response-usage metadata, reports turn triggers and sources in analytics and tracks TUI starts by app-server mode. Diagnostic-report uploads are made more resilient to slow networks. The changelog additionally records the removal of selected core test cases and redundant test coverage.

Interface refinements include opening the agents overview from an empty composer, refining hook-activity rendering and bounding background-terminal input previews. Taken together, the detailed changes make 0.153.0 a broad operational release covering interaction, state, approvals, plugins and internal infrastructure. The supplied source contains no benchmark results, deployment schedule, migration deadline or guarantee that its experimental context feature will become generally available.

Source: OpenAI Codex release notes for version 0.153.0.

Definition. OpenAI Codex 0.153.0 is an operational release focused on terminal history, reconnection, plugins, approvals, session state and context controls.