OpenAI Codex 0.159.0 adds an interrupt that steers a running agent, and protects .aws directories by default

Stopping a long-running coding agent to correct one line is expensive, because the next turn starts from scratch. OpenAI published Codex 0.159.0 on 29 September 2026 with an opt-in setting, instant_interrupt, that lets new input steer the run while the model is still responding or a code-mode call is still in flight.

What the interrupt does

The release notes name instant_interrupt as an opt-in feature that lets fresh input change direction during a model response or during a long-running code-mode call. On earlier versions, sending new text mid-turn either queued behind the current one or ended it, so a small correction cost a full restart. The setting has to be turned on, which is a signal that the vendor is treating this as a change in behaviour rather than a bug fix.

The release also lets a user scroll the transcript while deciding whether to implement a plan. Previously the choice sat on top of a static view. A new sessions welcome screen adds consistent headers and occasional tips during and after turns, and the warnings viewer now dismisses reviewed warnings when it closes, with k on a warning to keep it for later.

The security change to read

Approved commands now retain explicit filesystem denials, and .aws directories are protected by default under writable roots. On earlier versions an approved command inherited its writable-root list without carrying the denial rules that came with the original approval, which meant a shell command allowed to write inside a project could still touch an .aws folder that sat under the same root. Under 0.159, a Codex install that treats a home directory or a project directory as writable will refuse to write inside a .aws folder there unless the policy explicitly names it.

A separate fix restores macOS TLS access in network-enabled sandboxes and in remote environments that require proxy access.

What else moved

Windows launches no longer spawn stray console windows for MCP servers, code-mode hosts and piped commands, and a restrictive launcher can fall back to an embedded mode. Copying a transcript selection now preserves Markdown tables, formatting and significant whitespace, and more terminals copy automatically on selection. Blank sessions retain their drafts when switching tasks, and threads can be archived and listed before their first turn. Local ChatGPT sign-in opens the browser reliably, and onboarding provides a shortcut to copy the login link. Native Mermaid rendering supports more flowchart edges, labels and node groups. App-server clients can paginate thread history from a specific item.

Two things are gone. The release removes automatic follow-up prompt suggestions and the tui.prompt_suggestions setting that controlled them, and it removes the bundled plugin-creator skill.

What it changes for a team on Codex

Three practical shifts. Anyone who has watched a long code-mode call go the wrong way now has a switch to correct it without restarting: turn on instant_interrupt, send the correction, keep the turn. A team that runs Codex against a home directory or a broad project root can rely on .aws being off limits without writing that rule themselves. And a session that relied on the automatic prompt suggestions to nudge the next turn has to replace that habit, because both the suggestions and the setting are gone.

Source

Primary source: openai/codex release rust-v0.159.0 on GitHub, 29 September 2026.