Codex updates

Radar Ultra

Releases: 217

0.136.0

CLIGitHub

AI summary

  • You can archive sessions from the TUI or CLI. Archived sessions can’t be resumed or forked until you restore them.
  • Remote execution setup can register an API key for approved OpenAI hosts. Remote-control websockets now use short-lived server tokens instead of ChatGPT access tokens.
  • TUI Markdown links remain clickable, and cramped tables are shown as readable key/value records while keeping their link targets.
  • Windows admins can try an alpha elevated sandbox setup command; requirements can also specify allowed Windows sandbox implementations.
  • Several reliability and safety fixes improve token refresh handling, command protections, sandbox cleanup, resumed-session history, and app-server behavior.

Written by AI from the official notes. The original prevails.

26.601

AppOfficial notes

AI summary

  • You can now set where terminal tabs open: in General settings, choose the bottom or right panel for the terminal shortcut and environment actions. This choice applies when the bottom panel is enabled.
  • The release also includes additional performance improvements and bug fixes, with no further details provided.
  • The release also includes additional performance improvements and bug fixes, with no further details provided.

Written by AI from the official notes. The original prevails.

0.135.0

CLIGitHub

AI summary

  • Install scripts can now run without prompts when CODEX_NON_INTERACTIVE=1 is set.
  • /status shows remote connection details and the server version when the TUI is connected remotely.
  • codex doctor provides more environment, Git, terminal, app-server, and thread diagnostics for support cases.
  • The TUI has more readable Markdown tables and lists, and improved output stability on macOS and Zellij.
  • Other improvements include named permission profiles, expanded Vim editing, and slash-command completion that preserves draft text for commands accepting inline arguments.

Written by AI from the official notes. The original prevails.

26.527

AppOfficial notes

AI summary

  • Codex can now operate Windows desktop apps by viewing, clicking, and typing in the foreground.
  • You can start Codex work on a Windows device from ChatGPT on iOS or Android, or from Codex on a Mac, and check its progress remotely.
  • The Profile section now displays your profile details, usage statistics, and token activity.
  • You can search past Codex app threads using conversation content and Git branch names; Chrome context capture for Google Docs, Sheets, and Slides tabs has also been improved.
  • Other updates include coordination for local projects and worktrees, stable identicons for background subagents, and keyboard shortcut settings with keypress search and a reset-all action.

Written by AI from the official notes. The original prevails.

0.134.0

CLIGitHub

AI summary

  • You can now search your local conversation history, including case-insensitive matches, and see previews of results.
  • The CLI and related permission and sandbox flows now use --profile as the primary way to select a profile. Legacy profile configurations are rejected, with migration guidance provided.
  • MCP setup now supports targeting environment variables to individual servers and OAuth for streamable HTTP servers. Read-only MCP tools that advertise readOnlyHint can run concurrently.
  • Remote connections and operations are more resilient: Codex reconnects stale exec-server connections and retries certain remote actions after authentication recovery or stream failures.
  • Usage-limit errors now show messages specific to the affected workspace when credit or spending caps are reached.

Written by AI from the official notes. The original prevails.

0.133.0

CLIGitHub

AI summary

  • Goals are now enabled by default and track progress across active turns using dedicated storage.
  • codex remote-control now runs in the foreground, waits until ready, and reports its status. You can still use the separate start and stop commands for daemon-style use.
  • Permission profiles gained listing and inheritance support, managed requirements, runtime refresh, and stronger Windows sandbox integration.
  • Plugin listings now show more information, including installed versions and marketplace roots; remote collections are also supported.
  • Several reliability fixes improve working-directory selection, instruction loading, terminal cleanup, and app-server startup and shutdown. Code-mode output is also preserved unless you explicitly request an output limit.

Written by AI from the official notes. The original prevails.

26.519

AppOfficial notes

AI summary

  • On macOS, press both Command keys to send Codex a screenshot and available text from the frontmost app window, so it can use that context without you copying or describing it.
  • Goal mode is now generally available in the Codex app, IDE extension, and CLI. It lets Codex work toward an objective for hours or even days.
  • Codex can use desktop apps while your Mac is locked, including remotely through Codex Mobile. Locked use is limited to active, trusted computer-use turns and includes safeguards such as short-lived authorization and relocking after local input.
  • Browser tools can download and extract page images faster, extract structured data more effectively, and find information using a read-only JavaScript sandbox. Browser-use reliability has also been improved, including fixes for Windows and plugin availability in non-geo-blocked regions.
  • ChatGPT Business teams can share reusable plugin bundles through marketplace sources; Enterprise support is coming soon. In-app browser annotations can also adjust styling such as font size, colors, and spacing.

Written by AI from the official notes. The original prevails.

0.132.0

CLIGitHub

AI summary

  • The Python SDK adds built-in sign-in options, including API keys and ChatGPT browser or device-code login, plus account inspection and logout.
  • You can resume codex exec sessions with --output-schema, keeping session context while requiring structured JSON output.
  • Automations stop continuing a goal when they reach usage limits or hit the same blocker repeatedly, helping avoid unnecessary token use.
  • Remote sessions now keep websocket connections alive and show repository-relative diff paths rather than /tmp/... paths.
  • TUI startup is faster because terminal capability checks run in a batch instead of one after another.

Written by AI from the official notes. The original prevails.

0.131.0

CLIGitHub

AI summary

  • Codex adds codex doctor, which gathers support-ready diagnostics about runtime, sign-in, terminal, network, configuration, and local state.
  • Remote workflows gain codex remote-control, runtime enable/disable and status APIs, and support for registry-backed or configured remote environments.
  • The TUI gains richer session controls and information, including service-tier commands, token usage, permission and approval modes, workspace roots, and responsive Markdown tables.
  • A single @ picker can search files, directories, plugins, and skills.
  • Plugin workflows add marketplace commands, version-aware sharing, share checkout, clearer workspace groupings, and plugin hooks enabled by default.

Written by AI from the official notes. The original prevails.

26.513

AppOfficial notes

AI summary

  • You can use Codex through the ChatGPT mobile app by connecting it to a Mac running the Codex app. Codex runs on that Mac, so your phone can access its projects, files, credentials, plugins, skills, and configuration.
  • The launch also includes general availability for Hooks, Codex access tokens for trusted automation, and setup guidance for Enterprise admins.
  • Remote-connection guidance covers mobile setup, choosing a host, what is available from the connected machine, and SSH hosts.

Written by AI from the official notes. The original prevails.

0.130.0

CLIGitHub

AI summary

  • Plugin details now show bundled hooks, and sharing includes link metadata and controls for discoverability.
  • A new codex remote-control command provides a simpler way to start a headless app-server that can be controlled remotely.
  • App-server clients can page through large threads and choose whether to load no turn items, summaries, or full items.
  • Bedrock authentication can use AWS console-login credentials from aws login profiles.
  • Several reliability fixes improve live config updates, turn diffs after patch operations, thread handling, and remote compaction behavior.

Written by AI from the official notes. The original prevails.

26.506

AppOfficial notes

AI summary

  • You can now review hook trust in the app, and reach Hooks settings even before hooks are fully set up.
  • Your unfinished message edits are preserved when you switch threads.
  • Several desktop workflow issues were fixed, including pasting in the Windows terminal, opening modified external links outside the in-app browser, and keeping feedback commands with the right thread.
  • Loading and panel behavior was improved: resuming threads shows model loading, unavailable model controls stay hidden during loading, and summary panels received layout and hover fixes.
  • Other fixes restore tooltip-wrapped dropdowns, keep Computer Use settings visible when the feature is uninstalled, and disable problematic extension hover panels.

Written by AI from the official notes. The original prevails.

0.129.0

CLIGitHub

AI summary

  • The TUI is easier to use and resume: it adds Vim-style editing, a redesigned resume/fork picker, raw scrollback, /ide context injection, and workspace-aware /diff.
  • Plugin and hook workflows gain more controls: plugins can be shared and managed in more ways, while hooks can be browsed, toggled, and run around compaction. Some Codex Apps authentication and eligible MCP prompts can also appear through TUI or Guardian flows.
  • Experimental goals are easier to find, stay paused when you resume unless you opt back in, and show clearer validation and duration information.
  • Several everyday TUI issues were fixed, including copying in tmux, modified-key handling, Windows typing and paste delays, and preserving drafts through clear and editor workflows.
  • Reliability and security fixes improve Linux and Windows sandbox behavior and address issues including proxy-based login, unsafe project configuration, and unbounded MCP or hook output.

Written by AI from the official notes. The original prevails.

26.429

AppOfficial notes

AI summary

  • Dictation now includes cleanup and a configurable dictionary for names, file paths, and code symbols.
  • The image viewer now has zoom and download controls.
  • Pull request creation and recovery are more reliable: newly created pull request state is preserved, more app-server failures are classified as requiring a restart, and exhausted remote reconnect loops stop.
  • Voice and dictation errors now give clearer messages for microphone, connection, and quota failures.
  • In-app browser comment markers now stay aligned while scrolling, zooming, or changing responsive layouts.

Written by AI from the official notes. The original prevails.

0.128.0

CLIGitHub

AI summary

  • You can create and manage saved /goal workflows, including pausing and resuming them. They can continue across runtime sessions.
  • Codex adds plugin marketplace installation and remote bundle caching, along with external-agent session import, including background imports.
  • Permission profiles now have built-in defaults and can be selected from the sandbox CLI; clients can also see the active profile. The --full-auto option is deprecated in favor of explicit permission profiles and trust flows.
  • Several resume and interruption problems are fixed, including hangs from stale interrupts, restoring saved provider settings, handling large remote resume responses, and slow filtered resume lists.
  • The TUI gains configurable keymaps and fixes for terminal resizing, command popups, keyboard cleanup, and status updates. You can also edit /statusline and /title during an active turn.

Written by AI from the official notes. The original prevails.

0.125.0

CLIGitHub

AI summary

  • App-server integrations gained more connection and workflow options, including Unix sockets, pagination-friendly resume and fork, sticky environments, and remote thread configuration and storage.
  • Plugin management can now install remote plugins and upgrade configured marketplaces.
  • Permission profiles now carry across more parts of Codex, including TUI sessions, user turns, MCP sandbox state, shell escalation, and app-server APIs.
  • The app-server now respects project configurations explicitly marked untrusted instead of automatically saving them as trusted.
  • codex exec --json now includes reasoning-token usage for programs that consume its output.

Written by AI from the official notes. The original prevails.

26.423

AppOfficial notes

AI summary

  • Search in long review files now reliably jumps to matches that are off-screen when you move to the next or previous result.
  • Embedded MCP app panels no longer restart or lose their state during fullscreen changes and thread reloads.
  • Several desktop issues were fixed, including tray crashes when the local connection is missing, duplicate macOS fullscreen menu entries, and global dictation hotkeys failing on older macOS versions.
  • A tooltip on realtime delegation messages clarifies that Codex uses the surrounding voice conversation as context.

Written by AI from the official notes. The original prevails.

0.124.0

CLIGitHub

AI summary

  • Eligible ChatGPT plans now use the Fast service tier by default, unless you opt out.
  • App-server sessions can choose an environment and working directory for each turn, supporting more precise targeting in multi-workspace and remote setups.
  • Hooks are stable and can be configured in config.toml and managed requirements.toml. They can also observe MCP tools, apply_patch, and long-running Bash sessions.
  • Codex now supports Amazon Bedrock for OpenAI-compatible providers, including AWS SigV4 signing and AWS credential-based authentication.
  • The TUI has shortcuts to lower or raise reasoning (Alt+, and Alt+.). When you accept a model upgrade, reasoning resets to the new model’s default.

Written by AI from the official notes. The original prevails.

0.123.0

CLIGitHub

AI summary

  • Codex now includes Amazon Bedrock as a model provider, with support for choosing an AWS profile. Its bundled model information was also refreshed, including the current gpt-5.4 default.
  • Use /mcp verbose to see full MCP server diagnostics, resources, and resource templates; plain /mcp remains the faster option.
  • Several reliability fixes: /copy now copies the latest visible response after a rollback, follow-up text entered during a manual shell command is queued, and Unicode or dead-key input works in VS Code WSL terminals.
  • Plugin MCP configuration now accepts both mcpServers and top-level server maps in .mcp.json.
  • Remote environments can now require host-specific remote_sandbox_config settings.

Written by AI from the official notes. The original prevails.

26.422

AppOfficial notes

AI summary

  • GPT-5.5 is available in Codex and is recommended for most tasks when it appears in your model picker. If it isn’t available yet, update Codex; during rollout, continue using GPT-5.4.
  • In the Codex app, Codex can use the in-app browser with local development servers and file-backed pages—for example, to inspect a rendered interface, reproduce a visual bug, or check a local fix. Browser use runs through the bundled Browser plugin, whose settings let you manage the plugin and review allowed or blocked websites.
  • When configured, an automatic reviewer can check eligible approval prompts before the requested action runs. The app shows the review status and risk level so you can decide what to do next.
  • The app fixes duplicate queued follow-up messages across windows, chats reverting to outdated model and thinking settings, and edits to the same file no longer staying grouped after a chat is reopened.

Written by AI from the official notes. The original prevails.

Radar Ultra
An unofficial tool, not affiliated with OpenAI. Compiled from public posts; the original posts are authoritative.