Codex CLI 0.159 adds instant interrupt, so you can steer a runaway agent mid-thought

OpenAI shipped Codex CLI 0.159.0 (tagged rust-v0.159.0) on Tuesday, and the headline feature is aimed squarely at the control problem in agentic coding: an opt-in instant_interrupt that lets new input steer Codex during model responses or long-running code-mode calls, according to the release notes.

The wording in the release notes matters. "Steer" means your new input gets folded into an in-flight turn; it does not mean the CLI safely cancels every operation, undoes filesystem changes, or rolls back a half-executed plan. You still need Git checkpoints, permission boundaries, diff review, and sandbox policy for that. It is an interactive-control upgrade, not a safety guarantee.

The rest of the release is a tidy developer-experience package: a compact welcome screen with consistent headers and occasional tips, a warnings viewer that dismisses reviewed warnings when closed (press k to keep one), transcript scrolling while you decide whether to implement a plan, native Mermaid rendering with more flowchart edges, labels, and node groups, thread-history pagination for app-server clients, Windows launch fixes that avoid stray console windows, transcript-copy improvements that preserve Markdown tables and formatting, and draft recovery so blank sessions keep your drafts when you switch tasks.

On the safety side, approved commands now retain explicit filesystem denials, .aws directories are protected by default under writable roots, and macOS TLS access is fixed in network-enabled sandboxes and remote environments that need proxy access. Two removals worth noting: automatic follow-up prompt suggestions (and the tui.prompt_suggestions setting) are gone, and the bundled plugin-creator skill was removed.

Cadence note: this is the second Codex CLI release in two days, following 0.158.0 on September 28. OpenAI's Rust CLI is clearly in a rapid iteration phase, and 0.159 is the kind of release that makes the daily driver feel noticeably better without asking you to relearn anything.

The control-surface race

The broader significance is that every major coding-agent vendor is now competing on interruptibility, not just capability. Cursor shipped checkpointing and rollback earlier this year; Claude Code has had checkpoint-based rewind for months. OpenAI is late to explicit mid-turn steering, and this first version is deliberately conservative: opt-in, with no promise of undo. That honesty is the right call. The failure mode everyone fears is not a slow agent but an agent that keeps acting while you scramble for the stop button, and a steering input that lands mid-turn is a much safer mental model than a cancel button that might not fully cancel.

Worth remembering: this release lands four days after Codex's 56-minute full outage on September 25, when users were told their API keys were invalid. Tooling that misleads you at the front door and agents you cannot redirect in the middle are two faces of the same trust problem. Releases like 0.159, boring and practical, are how vendors fix the second half.