Codex CLI 0.167.0 lands an agent command center and Vertex AI setup

Anthropic's message threads also cut Claude Code latency ~35%, OpenAI's AI wrote 948 Linux root exploits, and Hugging Face pins model behavior in CI.

Nowline NOV 7 12:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • The agent command center gets unread flags

    `codex agents --no-daemon` now opens an agent hub that marks unread results, so you can fan out parallel runs and come straight back to the ones that finished. The session picker switches between connected backends mid-flow and keeps your drafts when you do.

  • Point Codex at Google Vertex AI

    A new `gcp_setup_wizard` setting walks you through Vertex AI using Application Default Credentials or a credential file, and a “Sign in another way” menu now collects API-key and cloud-provider logins in one place. You can also opt into system proxy settings from `/experimental` on macOS or Windows.

  • Claude message threads cut Claude Code latency

    Anthropic's new message-threads beta lets later API calls send only the new turn while the server keeps the conversation, trimming median time-to-first-token ~35% on Claude Code requests carrying 500K-1M tokens. Token counts and per-token rates are unchanged, and a thread expires after 5 idle minutes, so keep your own copy.

  • OpenAI's cyber model found 948 ways to root Linux

    An unreleased OpenAI cyber model dynamically reproduced 1,724 kernel vulnerabilities and wrote 948 working local-privilege-escalation exploits, at roughly $1,280 per vuln in API credits. The patches are candidates, not shipped fixes — a preview of what agentic bug-hunting against your dependencies now costs.

  • Hugging Face adds CI contracts to Transformers

    Model Integration Contracts pair a tiny model fixture, a runnable usage example, and reviewed expected outputs so a model's behavior is pinned in CI. Replaying one against history already caught a UMT5 masking bug that let the decoder peek at future tokens during training.