OpenAI
Codex CLI
1 day since its last release · 0.167.2 ·
Codex is an AI coding tool from OpenAI. Its latest release is 0.167.2, released on 11 Nov 2026. It shipped 7 releases in the last 30 days, and we have tracked 209 since 16 Apr 2025.
- Latest
- 0.167.2
- 11 Nov 2026
- Last 30 days
- 7
- releases
- Usual gap
- ~2d
- between releases
- Tracked
- 209
- since 16 Apr 2025
- Homepage
- github.com/openai/codex
Releases per month
last 12 months- 11 Dec
- 14 Jan
- 13 Feb
- 10 Mar
- 8 Apr
- 7 May
- 11 Jun
- 11 Jul
- 8 Aug
- 19 Sep
- 9 Oct
- 3 Nov
Release history
209 releasesReleases 80
-
Incremental tool updates now on by default for Vertex AI
- Tool updates are now sent incrementally by default when using Google Cloud Vertex AI
- Opt out by setting incrementaltoolslegacygcp = false under [features] in config.toml
-
Agent command center, backend switching and image previews in the TUI
- Open the agent command center with codex agents --no-daemon, and the command center and session picker now mark unread results
- Switch between connected backends from the session picker without losing your drafts
- Preview local image attachments while composing on supported terminals, and open viewed images from clickable transcript paths
- Google Cloud Vertex AI setup via the gcpsetupwizard option, with API key and cloud sign-in grouped under Sign in another way
- Saved chats stay visible after switching model providers, and Ctrl+L clears the screen while keeping earlier messages available through Ctrl+T
-
Keyboard-driven transcript selection and search in the TUI
- Transcript can now be selected from the keyboard with word, line and page motions, built-in search and configurable shortcuts
- GPT-6.1 Terra added to the bundled, Vertex AI and Amazon Bedrock model catalogs
- Safety blocks now come with clearer explanations and safer follow-up prompts you can pick from
- New codex queue --resume resumes unloaded threads, and codex doctor --network-only runs focused network diagnostics
- Markdown now renders consistently in async questions, choices, answers and transcript replies
-
Built-in Google Cloud Vertex AI provider and GPT-6.1 Terra model
- Added a built-in Google Cloud Vertex AI provider, set with modelprovider = "gcp"
- GPT-6.1 Terra added to the model catalogs
- Reasoning effort can now be changed mid-conversation by default on supported models
- Memory v2 is on by default for fresh profiles and can import memories from external agents
- Session picker gains a repository filter, and the agent command center groups sessions by directory across linked worktrees
-
Reusable prompt presets and inline /model and /permissions settings
- Save reusable prompts and slash commands in ten numbered /presets slots, with a keyboard menu for quick access
- Set model, reasoning effort, speed and permission profiles directly with inline /model and /permissions arguments
- Image previews in the fullscreen transcript on terminals that support Kitty or Sixel graphics
- Enterprise MCP connections can authenticate with Codex account grants or a dedicated sign-in, and credentials are masked in remote Bash and Zsh sessions
- Amazon Bedrock setup wizard is now on by default, plus fixes for compaction failures, session resume and Windows proxy authentication
-
Searchable settings panel, live steering mid-response and a leader key
- New /config panel searches preferences across categories, applies supported changes immediately and flags which ones need a restart
- Send a follow-up while Codex is responding to steer it, and interrupting a turn now also stops active Code Mode cells
- Configurable leader key in /keymap, defaulting to Ctrl+X, with voice mute moved to F9
- Independent read-only tools such as skill, memory and history lookups can now run in parallel
- Daybreak is on by default for eligible accounts, and codex exec can select it with an OpenAI API key when the server offers access
-
Managed Git worktree tools and task pinning in Command Center
- New tools create and list managed Git worktrees from trusted local projects when the worktrees feature is enabled
- Pin tasks in the agent Command Center with p and keep them in a shared Pinned group
- Copy transcript blocks with /copy, copy selections with Ctrl+Insert and adjust mouse-wheel scroll speed
- URLs are now clickable in approvals, questions, MCP prompts, warnings and banners, even when they wrap
- Custom Responses-compatible providers can enable live web access and remote compaction; applypatch now keeps CRLF line endings
-
GPT-6.1 Sol becomes the default model, plus opt-in Daybreak mode
- GPT-6.1 Sol is now the default model in the bundled and Amazon Bedrock catalogs
- Amazon Bedrock adds multi-agent V2 and Ultra reasoning on compatible models, and Bedrock Mantle accepts AWS GovCloud regions
- Daybreak is now opt-in via --enable clidaybreak, and Cyber access programs can be chosen per turn in codex exec and the TypeScript SDK
- Sign in to MCP servers from a running session with /mcp login <name>
- Voice conversations let you pick your microphone, speaker and input channels
-
Projectless sessions, task history paging and X11 middle-click paste
- Start sessions outside a project using workspace defaults when policy permits, with saved permissions restored on resume
- Browse older tasks in the agent command center with a keyboard-accessible Show more action
- Select transcript text and middle-click paste in fullscreen mode on supported local Linux X11 terminals
- Opt-in Guardian reviews can retrieve earlier user instructions and include context from agent handoffs
- Fixes for queued messages after reconnecting, Windows sandbox issues, SQLite stalls and provider model catalogs
-
Optional account security setup reminders for ChatGPT sign-ins
- Eligible local sessions signed in with ChatGPT can now show optional reminders to finish setting up account security
-
GPT-6.1 Sol becomes the default model, including on Amazon Bedrock
- GPT-6.1 Sol is now the default model in the bundled model catalog
- GPT-6.1 Sol added to the Amazon Bedrock Mantle and Runtime catalogs
-
Opt-in instant interrupt lets new input steer Codex mid-response
- Opt-in instantinterrupt lets new input steer Codex during model responses or long-running code-mode calls
- New sessions get a compact welcome screen, consistent headers and occasional tips during and after turns
- Copying transcript selections keeps Markdown tables, formatting and whitespace, and more terminals copy on selection
- Windows launches no longer open stray console windows for MCP servers, code-mode hosts and piped commands
- Automatic follow-up prompt suggestions, the tui.promptsuggestions setting and the bundled plugin-creator skill were removed
-
TUI copy-on-select and right-click paste, plus MCP OAuth secrets
- Fullscreen TUI adds configurable copy-on-select and right-click paste, and copied transcript selections keep their Markdown formatting
- Connect to MCP servers that need pre-registered OAuth client secrets, including via codex mcp add --oauth-client-secret
- Image generation and editing can request transparent backgrounds, and edits accept file-backed conversation images
- Terminal input approval is now on by default for commands with elevated permissions; exec-server WebSocket connections can use bearer tokens
- Fixes for Windows and Linux sandbox failures, unnecessary macOS approval prompts and Mermaid flowchart rendering
-
GPT-6 Sol and Luna models arrive, with fullscreen transcripts on by default
- Added GPT-6 Sol and Luna models, including Amazon Bedrock support and migration prompts for older models
- Fullscreen transcripts are now on by default, and Shift-click extends text selections
- Eligible interactive sessions now start a background server automatically, with recovery options when server settings are incompatible
- New f shortcut forks a conversation that is open in another app, keeping drafts and queued prompts
- /import now works in remote sessions and local background-server sessions
-
GPT-6 Sol and GPT-6 Luna added to the model picker
- GPT-6 Sol and GPT-6 Luna can now be chosen from the model picker
- The rate-limit switch prompt now recommends GPT-6 Luna
-
Optional fullscreen UI, voice on by default and a usage dashboard
- Optional fullscreen UI via /tui for the next launch, with transcript search, mouse selection and right-click copy
- Voice conversations now on by default, with an F8 toggle, a /voice settings picker and bundled audio runtimes for Linux and Windows
- New /usage analytics dashboard shows account usage, token totals and plugin and skill activity
- Agent command center can filter tasks by status and create worktree sessions, and worktrees are now on by default
- Six new terminal themes, Mermaid diagrams and display equations shown in responses, and sandbox isolation gaps closed
-
Experimental voice conversations and live reasoning summaries arrive in the TUI
- Experimental /voice conversations with live transcripts and microphone controls, enabled through /experimental on supported builds
- The TUI shows live reasoning summaries in the status row and completion timestamps after successful turns
- Agents overview adds task hiding, archiving and deletion, plus worktree ownership details and confirmed deletion of clean worktrees
- Touch ID verification for MCP requests in local TUI sessions on supported Macs
- Amazon Bedrock can get AWS credentials from configured commands, and saved threads recover after daemon restarts
-
GPT-6-Astra arrives, plus experimental worktree sessions
- GPT-6-Astra is now available in the model picker and Amazon Bedrock catalogs
- Experimental worktree support creates isolated checkouts for new or forked sessions via --worktree or /worktree, with a browser to resume them
- Answer Codex's questions inline while it keeps working, using suggested choices or custom text without losing your draft
- Windows sessions can share a background Codex server, with daemon lifecycle commands and managed updates
- The deprecated codex mcp-server command has been removed
-
Astra appears in the model picker and becomes the bundled default
- Astra now shows up in the bundled model picker
- Astra is now the bundled default model when no model is explicitly configured
- Astra's guidance now uses asynchronous questions only when that tool is available in the session
-
GPT-6-Astra now available through Amazon Bedrock
- GPT-6-Astra added to the Amazon Bedrock model picker for Mantle and Runtime global and US routes
- Fixed GPT-6-Astra guidance for async clarification questions to use the supported tool and accept text only
-
GPT-6-Astra can now be configured through the API
- GPT-6-Astra can be configured through the API without changing the default model
- GPT-6-Astra stays hidden from the model picker
- Backports the GPT-6-Astra model catalog to the 0.153 release line
-
Vim undo and redo, remote plugin marketplaces and earlier usage warnings
- Vim mode supports undo with u and redo with Ctrl+R, preserving full drafts including pasted content and attachments
- The plugin CLI can list, install and remove plugins from remote marketplaces
- Plus and Team users get an earlier warning when less than half their allowance remains in a roughly five-hour usage window
- TUI history shows complete patches, input sent to background terminals and individual completed commands, and tui.autorecap = false turns off automatic recaps
- TUI sessions reconnect after an app-server connection drops, keeping drafts and transcripts
-
Vim search in drafts and actionable rate-limit banners
- Vim mode adds / and ? search in drafts with highlighted matches and n/N navigation
- Rate-limit banners offer actions to check usage, manage credits, reset limits and manage plans
- MCP server names can use package-style characters, and individual MCP tools support an outputtokenlimit setting
- Planning tool is now off by default; enable it with tools.updateplan.enabled = true
- Fixes for automatic approval reviews, MCP tool availability, Windows sandbox PowerShell and cloud task credential safety
-
Extensions can now inspect or replace MCP tool results before the model sees them
- Extensions can now inspect or replace MCP tool results before they reach the model
- New configurable grace period for discovering tools from optional MCP servers
- Plugin catalogs combine per-repository configuration and flag invalid project marketplaces without hiding valid plugins
- Sandbox fixes: /cd can no longer weaken restrictions, and remote sandboxes respect the executor's home directory, OS and paths
- Tool availability and reasoning effort stay correct when switching or falling back between models
-
Mention and manage other Codex tasks from the terminal with @
- Reference other Codex tasks with @ mentions and ask agents to read, create or message tasks from the terminal
- /copy now offers a picker for full responses, individual code blocks and blockquotes
- Unnamed tasks get descriptive titles automatically, and /rename suggests an editable title based on the conversation
- Markdown links show as clickable labels in supported terminals, plus new shortcuts to cycle permission modes and Vim dot-repeat
- New Interrupt hooks, and untrusted projects can no longer supply project-level AGENTS.md instructions
-
New agents dashboard for managing tasks, plus message queueing to sessions
- New interactive codex agents dashboard to search, start, open, rename and stop tasks, with configurable shortcuts
- New codex queue command sends messages to existing local or remote sessions
- New /cd, /pwd and /cwd commands manage the working directory in TUI sessions
- codex doctor now diagnoses endpoint protection, network and proxy failures, desktop app state and update connectivity
- SDK adds exact CLI config overrides and max or ultra reasoning effort; Vim mode gains cw, c$, cc and character replace
-
Markdown conversation export, session forking and Amazon Bedrock support
- Export full TUI conversations to Markdown with /export, either to the clipboard or a new file
- Fork sessions with codex exec fork, and archive or restore sessions from the TUI resume picker
- Amazon Bedrock Runtime is now a built-in provider, with AWS profile, region and GPT-5.6 routing support
- Estimated thread credits or cost now appear in /status, status lines and terminal titles for eligible workspaces
- Hooks can run commands asynchronously and call MCP tools, and turns reconnect through temporary provider outages
-
Portable Agent Plugins, conversation sections and auto-approvals
- Install portable Agent Plugins and search local, personal, workspace and remote plugin catalogs
- Organize conversations into persistent, manually ordered sections and browse long transcripts incrementally
- New --approve-for-me flag enables automatically reviewed approvals
- Import Cursor-managed skills and sync imported Claude and Cursor conversations without duplicates
- Removed the deprecated codex exec --full-auto flag; use --sandbox workspace-write instead
-
Safer automatic-review defaults for cyber-capable models
- Automatic review now uses safer defaults when working with cyber-capable models
- The terminal interface now explains when and why permissions change
-
Named and pinned sessions, side conversations and plugin expansion
- Name new sessions with /new or /clear, pin important threads, and switch between side conversations without closing them
- Support for Agent Plugins manifests, workspace plugin publishing, and plugin marketplaces for Amazon Bedrock and Claude Code
- Fork threads with paginated history, including temporary forks that stay out of thread listings
- App-server can connect to remote Code Mode hosts over WebSocket, and compatible custom providers can use standalone web search
- Configured proxies are now honored across auth, plugin downloads, MCP, remote execution and LM Studio, plus terminal responsiveness fixes
-
Import from Cursor and Claude Code, audio input, and Bedrock login
- /import now migrates Cursor and Claude Code settings, MCP servers, plugins, sessions, commands and project-scoped memories
- Experimental Amazon Bedrock login with custom endpoints and authentication, using GPT-5.6 Sol as the default Bedrock model
- Audio inputs and tool outputs, including common local audio formats, plus streaming realtime V3 conversations
- Multi-agent V2 is now stable, with configurable sub-agent models, reasoning levels and concurrency
- Editing an earlier prompt now branches the conversation and keeps the original, and long conversations render faster in the terminal
-
Choose which usage-limit reset credit to redeem, plus new writes approval mode
- Usage-limit reset credits now show their type and expiration, and you can pick which one to redeem
- New writes app-approval mode allows declared read-only actions and prompts before writes
- MCP tools can request authentication interactively without an experimental opt-in
- Selecting Ultra reasoning warns when high multi-agent concurrency could increase usage quickly
- Fixes for Code Mode crashes on Intel macOS, Windows sandbox file deletion and pasted control sequences corrupting the TUI
-
Remote plugins on by default, plus system proxy support
- Remote plugins are now enabled by default, with richer catalog rows, npm marketplace sources and visible remote and local versions
- Authentication and Responses API traffic can go through macOS and Windows system proxies, including PAC and WPAD setups
- Added Amazon Bedrock GPT-5.6 Sol, Terra and Luna models, with first-class support for max reasoning effort
- New codex remote-control pair command generates manual pairing codes from a running daemon
- MCP tools now use tool search by default, alongside fixes for Windows ConPTY input, stale TUI prompts and exec server recovery
-
MCP tools now use tool search by default, plus macOS proxy support
- MCP tools now use tool search by default when supported, while staying compatible with older models and providers
- macOS authentication can honor system proxy, PAC and WPAD settings when respectsystemproxy is enabled
- Plugins can provide dedicated dark-mode logos through local manifests and remote catalogs
- PowerShell commands with regions the safety classifier cannot inspect now require approval
- Expired Amazon Bedrock credentials now give actionable recovery guidance instead of a generic error
-
Opt-in Windows system proxy support for sign-in
- Authentication can now use the Windows system proxy settings as an opt-in feature
- Supports PAC files, WPAD auto-discovery, static proxies and proxy bypass rules
-
Redeem usage-limit reset credits, token budgets and plugin recommendations
- /usage can now show and redeem earned usage-limit reset credits
- /plugins groups remote plugins into OpenAI Curated, Workspace and Shared with me, and can recommend and install relevant plugins
- Configurable token budgets track usage across agent threads, send remaining-budget reminders and stop turns when the budget runs out
- New indexed web-search mode allows live searches but limits direct page access to server-approved URLs
- Fixes Linux TUI rendering after Ctrl+Z suspend and keeps exec-server and MCP sessions alive through brief disconnects
-
Encrypted remote executors and cross-platform remote execution
- Remote executors now connect over authenticated, end-to-end encrypted Noise relay channels
- Cross-platform remote execution keeps the executor's own working directories, shells and filesystem permission paths
- Selected executor plugins can start their stdio MCP servers per thread, and plugin discovery adds a created-by-me marketplace
- TUI input prompts can resolve on their own after inactivity, with a countdown that pauses when you interact
- Faster, leaner large tool-heavy sessions, plus fixes for Windows sandbox credentials, hook trust in codex exec and enterprise proxy TLS
-
Usage views, Claude Code import, session deletion and Bedrock auth
- New /usage views show daily, weekly and cumulative account token activity
- New /import brings setup, project configuration and recent chats over from Claude Code
- Sessions can be permanently deleted with codex delete or /delete, with confirmation safeguards
- Adds managed Amazon Bedrock API-key auth and encrypted local storage for CLI and MCP OAuth credentials
- Experimental /realtime voice controls removed; corrupted SQLite state databases now rebuild automatically
-
Code mode gains direct web search; richer MCP tool schema support
- Code mode can call standalone web search directly, including from nested JavaScript tool calls, and get plaintext results
- Tool and connector input schemas now keep oneOf and allOf, improving compatibility with richer MCP tools
- codex doctor now reports editor and pager environment details, with raw values redacted in JSON output
- Plugin marketplace list JSON now shows each marketplace source, and plugin lists can load from the cached catalog
- Fixes for resume and fork with --last prompts, image edits targeting the wrong file, and sandbox proxy-only networking
-
Hand off CLI threads to Codex Desktop with the new /app command
- The /app command hands off the current CLI thread to Codex Desktop on macOS and native Windows
- Local image attachments and generated images now expose their saved file paths to the model for more reliable follow-up edits
- Reasoning effort selection adds fallback shortcuts and follows the effort levels and order the model advertises
- Plugin add, remove and marketplace commands support --json, and v2 personal access tokens now work for auth
- Goal workflow fixes, faster TUI startup and quicker resume --last on large session histories
-
Enterprise credit limits, cloud-managed config and remote-control pairing
- Enterprise and admin flows show monthly credit limits and can apply cloud-managed config bundles, including for EDU workspaces
- Remote-control clients can start pairing and list or revoke controller grants through new app-server v2 RPCs
- Hosted web and image tools work in more code-mode flows, and standalone web searches can run in parallel
- TUI adds F13-F24 keybindings, paste in searchable menus and a compact reasoning-only status item
- Cancelling a prompt before any output appears now restores the draft, attachments and collaboration mode
-
Session archiving, clickable terminal links and security hardening
- Archive sessions with /archive in the TUI or codex archive and codex unarchive in the CLI, so archived sessions can't be resumed or forked until restored
- TUI markdown keeps web links clickable, and cramped tables switch to readable key/value records
- Remote execution setup can now register with CODEXAPIKEY on approved OpenAI hosts
- Windows admins get an alpha codex sandbox setup --elevated provisioning command
- Security fixes stop /diff from running repository Git hooks, and deny read rules stay enforced for safe commands
-
Richer codex doctor diagnostics, Vim text objects and permission profiles
- codex doctor now reports environment, Git, terminal, app-server and thread inventory diagnostics for support cases
- Vim mode adds text-object editing, better word and line-end behavior, and a configurable interrupt-turn key
- /permissions understands named permission profiles and lists your custom profiles
- /status shows remote connection details and server version when connected over a remote transport
- Markdown tables and lists render more readably, and TUI output is more stable on macOS and Zellij
-
Search local conversation history and a unified --profile flag
- Search past local conversations with case-insensitive content matching and result previews
- --profile is now the main profile selector across CLI, TUI and sandbox; legacy profile configs are rejected with migration guidance
- MCP setup adds per-server environment targeting and OAuth options for streamable HTTP servers
- Read-only MCP tools marked with readOnlyHint can now run concurrently
- Fixes Windows TUI rendering corruption and improves remote connection reliability
-
Goals now on by default, plus a reworked remote-control command
- Goals are enabled by default with dedicated storage and track progress across active turns
- codex remote-control now runs in the foreground, waits for readiness and reports machine status, with start/stop kept for daemon use
- Permission profiles add list APIs, inheritance, managed requirements.toml support and stronger Windows sandbox integration
- Plugin list output now shows marketplaces, installed versions and remote collections
- Extensions can observe subagent start/stop, tool execution and turn metadata, alongside fixes for TUI working directory and Shift+Enter in plan mode
-
Python SDK gains built-in login and simpler text-only turn APIs
- Python SDK adds first-class auth: API key login, ChatGPT browser and device-code flows, account inspection and logout
- Python turns accept plain string input and return a richer TurnResult with items, timing and usage data
- codex exec resume now supports --output-schema for structured JSON output in resumed automations
- Faster TUI startup thanks to batched terminal capability probes
- Goal continuations now stop on usage limits or repeated blockers instead of looping and burning tokens
-
New codex doctor diagnostics, unified @ mentions and remote-control daemon
- New codex doctor command produces support-ready diagnostics for runtime, auth, terminal, network, config and local state
- @ mentions now search files, directories, plugins and skills in a single picker
- codex remote-control is now daemon-managed, with runtime enable/disable, status reads and configured remote environments
- Python SDK renamed to openai-codex with approval modes, concurrent turn routing and pinned generated types
- TUI status line shows blended token usage, permissions and approval mode, plus marketplace CLI commands for plugins
-
New remote-control command for headless sessions, plus plugin sharing
- New codex remote-control command starts a headless app-server that can be controlled remotely
- Plugin details now show bundled hooks, and plugin sharing adds link metadata and discoverability controls
- Bedrock auth can use AWS console-login credentials from aws login profiles
- App-server clients can page through large threads with unloaded, summary or full turn views
- Live app-server threads pick up config changes without a restart, and turn diffs stay accurate after apply-patch failures
-
Vim editing in the composer, redesigned session picker and hooks browser
- The composer now supports modal Vim editing, with a /vim command, a default-mode setting and Vim-specific keymaps
- Redesigned resume/fork picker, raw scrollback mode, /ide context injection and a workspace-aware /diff
- Hooks can be browsed and toggled from /hooks, run before or after compaction, and add PreToolUse context
- Plugins can be shared across a workspace with access controls, source filtering and marketplace upgrades or removal
- Status line can show theme colors plus PR and branch-change summaries, and Linux and Windows sandboxes are more reliable
-
Persistent /goal workflows you can pause, resume and clear
- New persisted /goal workflows with TUI controls to create, pause, resume and clear goals, plus app-server APIs and model tools
- Added codex update command, configurable TUI keymaps, plan-mode suggestions and action-required terminal titles
- Permission profiles expanded with built-in defaults and sandbox profile selection; --full-auto is now deprecated
- Plugins can be installed from a marketplace, uninstalled remotely and ship their own hooks; external agent sessions and configs can be imported
- Fixes for resume and interrupt hangs, TUI resize and layout, Windows sandbox edge cases, network proxy handling and Bedrock models
-
App-server gains Unix sockets, remote plugins and permission profiles
- App-server integrations support Unix socket transport, pagination-friendly resume/fork, sticky environments and remote thread config
- App-server can install remote plugins and upgrade configured marketplaces
- Permission profiles now carry across TUI sessions, user turns, MCP sandbox state, shell escalation and app-server APIs
- codex exec --json now reports reasoning-token usage
- Fixes for /review interrupt hangs, lost exec-server output, WebSocket disconnects and Windows sandbox startup
-
Stable hooks, Amazon Bedrock support and quick reasoning shortcuts
- Hooks are now stable, can be configured inline in config.toml and requirements.toml, and can observe MCP tools, applypatch and long-running Bash sessions
- Added first-class Amazon Bedrock support for OpenAI-compatible providers, with AWS SigV4 signing and AWS credential-based auth
- New TUI shortcuts Alt+, and Alt+. lower or raise reasoning, and accepted model upgrades now reset reasoning to the new model's default
- Eligible ChatGPT plans now default to the Fast service tier unless you opt out
- App-server sessions can manage multiple environments and pick an environment and working directory per turn
-
Built-in Amazon Bedrock provider and verbose MCP diagnostics
- Added a built-in amazon-bedrock model provider with configurable AWS profile support
- New /mcp verbose command shows full MCP server diagnostics, resources and resource templates
- Plugin MCP loading now accepts both mcpServers and top-level server maps in .mcp.json
- Realtime handoffs now send transcript deltas to background agents, which can choose to stay silent
- Fixed /copy after rollback, stuck Working states during shell commands and dead-key input in VS Code WSL terminals
-
Side conversations, fresh-context Plan Mode handoff and plugin upgrades
- The TUI can open /side conversations for quick side questions, and queued input now accepts slash commands and ! shell prompts while work runs
- Plan Mode can start implementation in a fresh context and shows context usage before you decide whether to carry the planning thread forward
- Plugin workflows add tabbed browsing, inline enable/disable toggles, marketplace removal and remote, cross-repo or local marketplace sources
- Filesystem permissions support deny-read glob policies, and codex exec can run isolated from user config and rules
- Tool discovery and image generation are now on by default, and codex app now opens or installs Desktop on Windows and Intel Macs
-
Plugin marketplace installs, Ctrl+R history search and memory controls
- New codex marketplace add command installs plugin marketplaces from GitHub, git URLs, local directories or marketplace.json URLs
- TUI prompt history gains Ctrl+R reverse search and local recall of accepted slash commands
- New TUI and app-server controls for memory mode, resetting and deleting memories, and memory-extension cleanup
- Expanded MCP support with MCP Apps tool calls, namespaced tool registration and opt-in parallel calls
- Secure devcontainer profile with bubblewrap support, plus Unix socket allowlists in the macOS sandbox
-
Realtime V2 streams background agent progress and queues follow-ups
- Realtime V2 streams background agent progress while work runs and queues follow-up responses until the active one completes
- TUI hook activity is easier to scan, with running hooks shown separately and completed output kept only when useful
- Custom TUI status lines can show the renamed thread title
- SessionStart hooks can tell sessions started by /clear apart from fresh starts and resumes
- Fixes for Windows and symlinked sandbox permissions and for codex --remote panics over secure websockets
-
Realtime voice moves to WebRTC by default, plus richer MCP support
- Realtime voice sessions now default to the v2 WebRTC path, with configurable transport, voice selection and native TUI media support
- MCP Apps and custom MCP servers gain resource reads, tool-call metadata, custom-server tool search, elicitations and file uploads
- Remote workflows add egress websocket transport, remote --cd forwarding and an experimental codex exec-server subcommand
- Ctrl+O copies the latest agent response, and /resume can jump straight to a session by ID or name
- Faster TUI startup, accurate /status rate limits, more configurable notifications and many resume and sandbox fixes
-
Proxy-only networking for Windows sandbox and stdin piping for codex exec
- Windows sandbox runs can enforce proxy-only networking with OS-level egress rules instead of relying on environment variables alone
- codex exec now accepts piped stdin input alongside a separate prompt on the command line
- Custom model providers can fetch and refresh short-lived bearer tokens dynamically instead of using only static credentials
- App-server clients can start ChatGPT sign-in with a device code flow when browser callback login is unreliable
- Project-local .codex files are now protected on first creation, and TUI regressions in /copy, /resume and /agent are fixed
-
Plugins become first-class, with a /plugins browser and installs
- Plugins are now a first-class workflow: Codex syncs product-scoped plugins at startup, and you can browse, install and remove them in /plugins
- Sub-agents use readable path-based addresses like /root/agenta, with structured inter-agent messaging and agent listing
- The app-server-backed TUI is now on by default and gains prompt history recall across sessions plus a /title terminal-title picker
- App-server clients can run ! shell commands, watch filesystem changes and connect to remote websocket servers with bearer-token auth
- Fixes for duplicated output, ChatGPT login, terminal restore on early exit and Linux and Windows sandboxing
-
Plugin install prompts, a new prompt hook and ChatGPT device sign-in
- Codex can now prompt you to install missing plugins or connectors, honor a suggestion allowlist and sync install or uninstall state remotely
- New userpromptsubmit hook lets you block or add to prompts before they run and before they enter history
- App-server TUI supports device-code ChatGPT sign-in during onboarding and can refresh existing ChatGPT tokens
- Realtime sessions start with recent thread context and are less likely to interrupt themselves during audio playback
- Fixed a first-turn stall, restored history for remote resume and fork, and improved Linux sandbox startup
-
Smart Approvals guardian, full-resolution images and app-server file APIs
- Smart Approvals can route review requests through a guardian subagent, cutting repeated setup on follow-up approvals
- Supported models can request full-resolution image inspection via viewimage and codex.emitImage
- The v2 app-server adds filesystem RPCs for reads, writes, copies, directories and path watching, plus a new Python SDK
- Realtime websocket sessions gain a transcription mode, v2 handoff through the codex tool and a unified [realtime] config
- Subagents inherit sandbox and network rules more reliably, and the TUI no longer stalls on exit after creating them
-
Experimental code mode and hooks engine, plus approval fixes
- Added an experimental code mode for more isolated coding workflows
- Added an experimental hooks engine with SessionStart and Stop hook events
- Approval flows now keep granted permissions across turns and applypatch honors them
- WebSocket app-server deployments now expose /readyz and /healthz health check endpoints
- New config switch to disable bundled system skills, and the $ mention picker now labels Skills, Apps and Plugins
-
Runtime permission requests, plugin marketplace and finer sandbox policies
- New built-in requestpermissions tool lets running turns ask for extra permissions, with approval prompts shown in the TUI
- Plugins gain curated marketplace discovery, richer list metadata, install-time auth checks and an uninstall endpoint
- New permission-profile config language with separate filesystem and network sandbox policies for more precise control
- Web search settings now accept full tool configuration such as filters and location, not just on/off
- Image generation saves output files to the current working directory; CLI releases now publish to winget
-
Mention plugins with @ to pull their tools and skills into chat
- New @plugin mentions let you reference a plugin in chat and automatically include its MCP, app and skill context
- The model picker in the TUI now shows the latest model catalog changes
- Skill permission profiles now merge into each turn's sandbox policy for zsh-fork runs, so privileges are added more safely
- JS REPL keeps previously set bindings after a failed cell and only accepts data: URLs for emitted images
- Sandbox fixes: Linux bubblewrap always unshares the user namespace, and macOS Seatbelt handles network and unix sockets more reliably
-
Fast mode on by default, with the current speed shown in the header
- Fast mode is now enabled by default, and the TUI header shows whether the session is running in Fast or Standard mode
- jsrepl can now dynamically import local .js and .mjs files to reuse workspace scripts
- Codex tells the model which plugins are enabled at session start, improving discovery of installed MCPs, apps and skills
- App-server v2 exposes MCP elicitation as a structured request/response flow, and clients get expanded image workflow support
- Resuming a thread now keeps its stored git context and keeps apps enabled
-
Plugin system for skills, MCP servers and app connectors
- New plugin system loads skills, MCP entries and app connectors from config or a local marketplace, with an install endpoint in the app server
- Multi-agent flow in the TUI gains approval prompts, enablement via /agent, ordinal nicknames and role-labeled handoffs
- Persisted /fast toggle in the TUI, plus app-server support for fast and flex service tiers
- Memories now support workspace-scoped writes, renamed settings and guardrails against saving stale or polluted facts
- Faster, more reliable sub-agents and a new direct Windows installer script
-
Fork threads into sub-agents, plus voice device picker
- Threads can now be forked into sub-agents to branch work without leaving the current conversation
- Realtime voice sessions let you choose and persist microphone and speaker devices
- Custom tools can return multimodal output such as images, not just plain text
- Memories are now configurable, with a new codex debug clear-memories command to fully reset them
- Fixes duplicate final responses in the terminal and restores pending approvals when resuming threads
-
Direct install script for macOS and Linux, plus memory upgrades
- New direct install script for macOS and Linux, published as a GitHub release asset
- jsrepl moved to /experimental with Node compatibility checks and a lower minimum Node version of 22.22.0
- requestuserinput now works in Default collaboration mode, not just Plan mode
- 5.3-codex now appears in the CLI model list for API users, and memory gains diff-based forgetting and usage-aware selection
- Fixed a zsh-fork path that could bypass sandbox filesystem restrictions, and added a ~1M-character input cap to prevent hangs
-
Syntax highlighting, theme picker and voice dictation come to the TUI
- TUI now syntax-highlights code blocks and diffs, with a /theme picker, live preview and theme-aware diff colors for light and dark terminals
- Experimental voice dictation: hold the spacebar to record and transcribe prompts (enable features.voicetranscription in config)
- Multi-agent upgrades: spawnagentsoncsv fans out work from a CSV with progress and ETA, plus sub-agent nicknames and a cleaner picker
- New /copy command copies the latest reply, while /clear and Ctrl-L clear the screen without losing thread context
- Codex can request extra sandbox permissions for a command, and specific approval prompt types can be auto-rejected
-
Websocket proxy support and thread archive notifications
- Network proxy now honours WSPROXY and WSSPROXY environment variables, including lowercase variants, for websocket traffic
- App-server v2 sends notifications when threads are archived or unarchived, so clients no longer need to poll
- Command approvals now carry distinct IDs, allowing multiple approvals within a single shell command
- Ctrl+C or Ctrl+D now cleanly exits the working-directory prompt when resuming or forking a session
- Fewer false-positive safety-check downgrades by reading the model from response headers instead of the body
-
Codex-managed commit co-author attribution and richer app listings
- Commit co-author attribution now uses a Codex-managed prepare-commit-msg hook, configurable via commandattribution (default, custom label or off)
- App listing responses now include app metadata, branding and labels so clients can render fuller app cards
- Removed the remotemodels feature flag to improve model selection reliability and performance
-
Unified permissions flow and structured network approval prompts
- More unified permissions flow, with clearer permissions history in the TUI and a slash command to grant sandbox read access to blocked directories
- Network approval prompts now show richer host and protocol context
- Multi-agent roles can now be customized via config, with collab settings renamed to multi-agent
- App-server adds fuzzy file search session-complete signaling and a model reroute notification for clients
- Fixes for remote image attachments on resume, a screen reader regression, thread resume and jsrepl stability
-
Experimental JavaScript REPL, memory commands and multiple rate limits
- Added an experimental, feature-gated JavaScript REPL runtime (jsrepl) that keeps state across tool calls
- New TUI memory slash commands /mupdate and /mdrop for managing memories
- Support for multiple simultaneous rate limits, shown in the TUI status surfaces
- Sandbox capabilities promoted on Linux and Windows, with a new configurable ReadOnlyAccess read-access policy
- Apps SDK apps now allowed in ChatGPT connectors, and app-server websocket transport is back, alongside websocket stability fixes
-
Shell commands now run alongside active turns without interrupting them
- Direct shell commands no longer interrupt an in-flight turn and can run concurrently while a turn is active
- New /statusline command lets you choose interactively which metadata appears in the TUI footer
- Resume picker can sort sessions by creation time or last-updated time
- Image attachments now accept GIF and WebP, and admins can restrict web search modes and network access via requirements.toml
- Fixes for Windows sign-in startup, required MCP servers now fail fast, and a file-watcher bug that produced very large logs
-
GPT-5.3-Codex arrives and steer mode is now on by default
- Adds the new GPT-5.3-Codex model
- Steer mode is now stable and on by default: Enter sends right away during running tasks, Tab queues follow-up input
- Correct instructions are now applied when switching models mid-conversation or resuming with a different model
- Fixes a remote compaction mismatch that could cause context overflows
- Default assistant personality restored to Pragmatic
-
Remember MCP tool approvals per session and live skill reloading
- New Allow and remember option auto-approves repeated MCP and App tool calls for the rest of the session
- Skill file changes are now picked up live without restarting
- New /debug-config slash command in the TUI shows the effective configuration
- Logs can be redirected with a new configurable logdir setting, including via -c overrides
- Fixes for TUI picker jitter, the working status indicator and cloud requirements reliability
-
Async thread compaction API and unified exec on macOS and Linux
- Added an async thread/compact call to the v2 app-server API so clients can start compaction and track completion separately
- Enabled unifiedexec on all non-Windows platforms
- New codex.ratelimits event signals rate limits over websockets
- Config requirement values now show where they came from, helping debugging in views like /debug-config
- Esc in the requestuserinput overlay now closes notes instead of interrupting the session
-
Launch the desktop app from the CLI, load personal skills and run shell tools in parallel
- New codex app command on macOS opens Codex Desktop from the CLI and downloads it automatically if it is missing
- Personal skills now load from ~/.agents/skills, and new app-server APIs list and download public remote skills
- /plan now accepts inline prompt arguments and pasted images
- Shell tools can run in parallel, and shell runs now get CODEXTHREADID so scripts can detect the active thread
- Git commands that are destructive or can write no longer skip approval checks, and codex exec no longer hangs after an interrupt
-
Plan mode on by default, plus stable personality settings
- Plan mode is now enabled by default, with updated interaction guidance in the plan prompt
- Personality configuration is now stable: the default is friendly, the config key is personality, and existing settings migrate forward
- Skills can be loaded from .agents/skills, with clearer relative-path instructions and support for nested folders
- Console output now shows runtime metrics to make diagnostics easier
- Fixes for sidebar ordering after unarchiving a thread and for repeated conversation rules
-
Plan mode gets its own view, and smart approvals are on by default
- Plan mode now streams proposed plans into a dedicated TUI view, with a feature-gated /plan shortcut for switching modes
- Smart approvals are enabled by default, with explicit approval prompts for MCP tool calls
- New /apps command to browse connectors in the TUI, plus $ insertion for app prompts
- Optional SOCKS5 proxy listener with policy enforcement, and app-server support for external ChatGPT auth
- New SQLite-backed log database with thread-id filtering, retention and an improved logs client
Smaller updates 20
-
Patch release with no detailed changes listed
-
Fixes a multi-line question crash and background server startup failures
-
Fix for Windows environment variables in remote MCP servers
-
Stops console windows flashing on Windows
-
Patch release with no detailed changes listed
-
Fixes rejected requests by turning reasoning summaries off by default
-
Fixes the speed shown for the GPT-6-Astra Fast tier
-
Guardian approval review now respects Node REPL policies
-
Remote compaction now counts retained images toward token budget
-
Minor patch release with no detailed changes listed
-
GPT-5.6 Sol, Terra and Luna context windows corrected to 272K
-
Better dangerous-command detection with clearer denial reasons
-
Maintenance patch with no user-facing changes
-
Version-only release with no code changes
-
Rolls back a Guardian auto-review prompting regression
-
Fixes for standalone installs and code mode reliability
-
Stops full WebSocket request payloads from appearing in trace logs
-
Maintenance release with no user-facing changes
-
Maintenance-only patch with no user-facing changes
-
Memory pipeline fixes and stable model selection by prefix
Showing the newest 100 of 209.