OpenAI Codex CLI 0.163: Mid-Response Steering by Default
OpenAI shipped Codex CLI 0.163.0 to general availability on 15 October 2026, making mid-response steering the default - you can redirect the agent while it is still generating - plus a searchable /config panel, a leader-key shortcut scheme and parallel read-only lookups
Tool Interface
Interactive tool will be available soon
Features
- ✓ Mid-response steering is now the default: send a follow-up while the agent is still generating and it redirects immediately
- ✓ Interrupting a turn also stops the active Code Mode cells and their nested tool calls, so the blast radius is explicit
- ✓ A new searchable /config panel puts model, approval policy and keymaps in one place instead of hand-editing config files
- ✓ Read-only lookups now run side by side, so the reconnaissance-before-editing step costs far less waiting
- ✓ Plenty of daily friction fixed: DNS and proxy connectivity, duplicate Host headers, credential aliases, Windows paste, WSL clipboard and late-arriving output from remote commands
How to Use
- Upgrade first: bring codex to 0.163.0 with npm or your package manager, then confirm with codex --version
- Open /config once and settle the model, approval policy and keymap defaults rather than living with an old config
- Remove retired shell flags from your config, or unattended runs will fail on the very first step
- Before leaving the agent unattended, re-run one trusted command (pwd, a dry-run test, git status) and only then start long tasks
FAQ
What is the headline change in 0.163.0?
Mid-response steering went from an opt-in flag to the default. Previously you either waited while the agent generated or hard-interrupted it; now you can add direction while it is still writing and it follows immediately, which feels closer to pair programming than babysitting.
How is steering different from interrupting?
Steering continues the same turn in a new direction; interrupting terminates it. Per the release notes, interrupting a turn also stops active Code Mode cells and their nested tool calls, so the effect reaches execution work and not just the visible response.
What should I check after upgrading?
First remove retired shell flags from your config - that is the most common post-upgrade breakage. Then confirm approval policy and keymap match your habits. For long tasks the recommended pattern is to park the thread and resume later, with approvals coming back with it.
What can the /config panel change?
Common defaults such as model, approval policy and keymap are searchable and editable in the panel, so you no longer hand-edit config files. Because it is searchable it beats digging through docs when there are many options.
Anything better for Windows and WSL users?
On Windows both Ctrl+V and Shift+Insert paste text and images; under WSL repeated pastes reuse the clipboard reader and copy attempts time out instead of hanging; local commands return promptly when a child keeps an output pipe open, and remote commands preserve late-arriving output for the final poll.
What is it good for?
Terminal coding work that needs back-and-forth correction: long refactors, bulk file edits, writing tests. Steering drops the cost of saying "no, take the other path" to a single sentence - but still keep destructive operations in a session you are watching.