- OpenAI's Codex CLI 0.122.0 lands with a rebuilt plugin UX (tabs, toggles, remote marketplaces), a brand-new /side slash command, slash + shell prompts in queued input, and a Plan Mode that can launch execution in a clean context with usage preview.
TL;DR
OpenAI shipped Codex CLI 0.122.0 on April 20, 2026. Three headline shifts: the plugin system grew a real UI (tabs, inline toggles, marketplace add/remove, remote and cross-repo sources); the TUI learned /side for quick side conversations plus slash commands and shell prompts inside queued input; and Plan Mode can now start execution in a fresh context with a usage preview before you commit. Sandboxing also tightened with deny-read globs and trusted-workspace hooks.
What's new
- Plugin tabs + inline toggles. Plugin management moves from a flat list to a tabbed workflow. Enable/disable without opening a settings pane, remove from the marketplace in one step, and pull plugins from remote, cross-repo, or local marketplace sources.
/sideconversations. A new slash command opens a side thread for quick questions without polluting the main context. Ask "what does this regex do?" while the agent keeps building.- Slash + shell in queued input. While the agent is working, you can queue not just text, but slash commands and shell prompts. Stack
/planor a shell command mid-run instead of waiting. - Plan Mode, fresh context. Plan Mode can launch implementation in a clean context window rather than forcing the planning thread forward. A context-usage preview shows before you decide whether to carry the planning history.
- Standalone installer fixes.
codex appnow opens or installs Codex Desktop correctly on Windows and Intel Macs; installs are more self-contained. - Tighter sandboxing. New deny-read glob policies, managed requirement enforcement, platform sandbox hardening, and isolated
codex execruns. - Defaults changed. Tool discovery and image generation are now on by default, with higher-detail image handling.
Why it matters
The 0.122 release is about two long-standing pain points: context bloat and flow interruption. Plan Mode used to shove a bulky planning transcript into the execution context, eating tokens and hurting model focus. Branching into a fresh context with an upfront usage preview fixes that directly. On the flow side, /side plus queued slash/shell commands means you no longer stall the agent to ask a quick question or sneak in a status check. Plugins getting a real UI is the quiet story: once installing and toggling extensions is frictionless, a third-party ecosystem actually has room to grow.
Technical facts
- Version: 0.122.0 (stable, cut from 0.122.0-alpha.7).
- Released: April 20, 2026.
- Install: existing installer,
npm i -g @openai/codex, or standalone binary. - Desktop fix:
codex appon Windows and Intel Macs no longer double-installs or fails to open. - Sandbox: deny-read glob policies, per-run isolation for
codex exec, trusted-workspace requirement for project hooks. - Security: logout now revokes managed tokens (previously lingered).
- Bug fixes: stale app-server prompts clear when resolved by other clients; remote control startup tolerates missing ChatGPT auth; resumed threads replay token usage immediately so the context meter is correct on resume.
Comparison vs 0.121.x
| Area | 0.121.x | 0.122.0 |
|---|---|---|
| Plugin UI | Flat list | Tabs, inline toggles, marketplace add/remove |
| Plugin sources | Local / single source | Remote, cross-repo, local marketplaces |
| Side conversations | Not available | /side slash command |
| Queued input | Text only | Text + slash + shell prompts |
| Plan Mode context | Carried forward (heavy) | Fresh context option with usage preview |
| Sandbox controls | Basic allow/deny | Deny-read globs, trusted-workspace hooks |
Use cases
- Long planning sessions: design in Plan Mode, check the context-usage preview, launch execution in a clean context. Keeps the coder fast even when the plan is huge.
- Quick detours:
/sidemid-task for "what does this API return?" without breaking the main thread. - Team plugin distribution: point the marketplace at an internal Git repo so every engineer installs the same extension set with one click.
- Mid-run command stacking: queue a
/planor shellgit statuswhile a long edit finishes — no flow break. - Locked-down monorepos: deny-read globs plus trusted-workspace hooks let Codex run closer to sensitive files without trust drift.
Limitations & pricing
Codex CLI stays free and open-source; model usage still bills against your OpenAI API credits or ChatGPT Plus/Pro entitlements. One known Plan Mode edge case (GitHub issue #18496): security-redaction placeholder blocks can fail exact replacement while Plan still reports "ready to implement" — workaround is re-plan or edit manually. The Desktop installer fix covers Windows and Intel Macs; Linux standalone still follows the prior install path. And because the plugin marketplace now accepts remote sources with no central review step, trust becomes the user's responsibility — treat third-party plugin installs like any other supply-chain action.
What's next
Direction is clear from what 0.122 invests in: sandbox policies will keep hardening, the plugin ecosystem has room to grow now that the UX is in place, and the separation between planning and execution context is likely to deepen — expect the queued-input surface to keep expanding in upcoming minors. If you run Codex daily, the upgrade is a no-brainer: the /side + queued-input combo alone earns back its install time within an afternoon.
Nguồn: Codex Changelog on X, openai/codex releases, OpenAI Developers Codex changelog.
