OpenAI Codex CLI 0.167.0: Unread Agent Results, Draft-Preserving Backend Switching and Patch Review

What shipped, and when​

The GitHub release page shows version 0.167.0, tagged rust-v0.167.0, as a stable release. The release notes record the full changelog as rust-v0.166.0 to rust-v0.167.0. The notes give a publication time of 01:40 UTC on November 6. This was not a surprise drop. The release followed a run of alpha builds, with alpha.2 and alpha.3 appearing on November 1 and alpha.5 on November 4.

The notes include no benchmarks or productivity figures. Everything below describes behavior, not measured results.

Agent controls in the terminal​

• Command center: You open it with codex agents --no-daemon. The release notes say the command center and the session picker now flag unread results.
• Backend switching: The terminal session picker can switch between connected backends and keeps your drafts when you do.
• Image previews: On supported terminals, you can preview local image attachments while composing a message. Viewed local images open from clickable paths in the transcript.
• Inline patch review: Patch details can be expanded in the transcript while an approval prompt is open.
• Terminal fixes: Ctrl+L clears the display but keeps earlier messages reachable through Ctrl+T. Deleting text that exposes a leading ! no longer flips an ordinary draft out of chat mode.

An unread marker only says you haven't looked at an output yet. It says nothing about whether the output is correct or deserves your attention first. The same goes for inline patch expansion. It puts the proposed change next to the approve button, but it doesn't make the patch safer.

Provider changes: reading a chat is not resuming it​

Saved chats now stay visible after you switch model providers. If a chat can't resume, you can read its transcript and prepare an editable prompt to continue in a new chat.

Read the wording carefully. The notes don't say an incompatible session can carry on unchanged, and they don't say the prepared prompt is sent automatically. You get your context back. You don't get guaranteed execution continuity. Anyone who swaps providers, for example when a team changes cloud vendors, should expect to start a fresh chat and seed it from the old record.

Setup, proxy and policy changes​

• Sign-in menu: API-key and cloud-provider options now sit under "Sign in another way."
• Vertex AI wizard: You enable gcp_setup_wizard to get a Google Cloud Vertex AI setup flow. It works with Application Default Credentials or a credential file. It is off by default, so nothing changes until you turn it on. The release notes say little about permissions, regions or model availability, so check those against your own Google Cloud project.
• System proxy (Windows and macOS): You opt in from /experimental, and the release calls it experimental. Windows admins behind corporate proxies should test it in a pilot before relying on it. The release makes no compatibility promise for managed networks.
• Documentation: The app-server README now covers managed network constraints under [permissions.network]. It also covers prompt-user command rules that require fresh human approval. These are documentation updates. They don't tell you how any given organization has configured these controls.

Approval and accounting fixes​

• Network approvals: Prompts close when their underlying requests end. Concurrent requests to the same endpoint are prompted one at a time.
• Context compaction: Automatic compaction now accounts for retained output from failed responses and for responses with missing token-usage data. This concerns when compaction triggers. It doesn't enlarge the context window, and it doesn't promise that compaction keeps the details you care about.
• Realtime voice: In long Realtime V3 conversations, completed speech is saved to chat history as the transcript grows.
• Linux musl builds: They use fewer allocator arenas to cut idle memory after conversations unload. The notes give no measured saving.
• Packaging and docs: DotSlash manifests now reference the Cargo packages shipped with the release. The openai-docs skill now recommends the documented replacement when you migrate off a deprecated model.

What to test before you roll it out​

1. Switch providers on a test machine and confirm an older chat still appears in the resume picker and opens as a readable transcript.
2. Trigger an approval prompt and check that expanded patch details stay in view and your scroll position holds.
3. If you use Google Cloud, enable gcp_setup_wizard and try both the Application Default Credentials path and the credential-file path.
4. On Windows, try the experimental proxy setting from a network that uses a system proxy. Keep a rollback plan.

Bottom line​

This is a workflow-maintenance release. It targets specific failure modes: lost drafts, hidden chats, cramped approval prompts and overlapping network prompts. The release notes don't claim better code or faster delivery, and neither should anyone summarizing them. A comparison with Claude Code can only be about the scope of each tool's work surface. Nothing in the release supports a performance lead either way. Whether these changes actually save you time is something your own team will have to measure.