Summary
OpenAI Codex CLI is OpenAI's terminal-based coding agent, shipped as openai/codex under rust-v* release tags. The October 2026 releases (v0.161–v0.166, Oct 27–29) are dominated by three themes: making the agent steerable while it works (mid-turn steering, prompt presets), broadening its model/provider surface (Google Vertex AI, the GPT-6.1 model family), and hardening failure semantics (terminal provider errors, MCP cancellation, token accounting across resume/fork, URL host-check bypasses). This page distills that release window and the reference comparison against a peer agent (yoyo at 0.25.16) that highlights where Codex is ahead.
Key Points
- Mid-turn steering (v0.163) is the headline interaction change: a follow-up message can be sent while the agent is still responding, redirecting work without cancelling and restarting the turn.
- Reusable prompts arrived in v0.164 as ten
/presetsslots for saved prompts and slash commands;/modeland/permissionsnow take inline arguments. - Provider and memory changes in v0.165: a built-in Google Vertex AI provider (
modelprovider = gcp) and "memory v2" enabled by default. - Model lineup shift: GPT-6.1 Sol became the default in v0.161; GPT-6.1 Terra was added to the catalogs in v0.165/0.166.
- Fixes define expected agent behavior: model-not-found is terminal (no retry, provider message preserved), dropped MCP tool futures cancel the request, reasoning-token accounting survives resume/fork so compaction isn't triggered prematurely.
- Security fix: marketplace URLs containing userinfo with a backslash are rejected — a host-check bypass.
- Relative to a peer agent (yoyo 0.25.16), Codex's lead is in interaction and infrastructure — steering, Vertex, presets — not in model catalog coverage, where yoyo already ships Sol presets.
Concepts
- Mid-turn steering — injecting user input into an in-flight agent turn.