What happened
OpenAI’s GitHub release notes list Codex CLI 0.168.0 as released on November 13, 2026. The release adds a model-provider view under Sessions → Model providers, where users can browse providers and select the default for new chats. The notes also document an important persistence rule: open and resumed chats continue to use their saved model provider after the default for new chats has changed. In other words, changing a default affects future conversations rather than silently reconfiguring an existing one.
The terminal interface gains an opt-in workspace capability through features.tui_workspace. It displays a side conversation beside the main chat, with a draggable divider; users can also press leader-z to zoom the focused pane. OpenAI describes it as feature-gated rather than default behavior. The release also fixes steering-message delivery during final transcript saving and when returning to a long-running chat, and it changes how replies interrupted by tool calls are rendered and replayed.
For media workflows, Codex can attach recordings when it is used with an audio-capable model. According to the release notes, recordings confirmed as rejected are skipped on later turns, while clip duration is used to retain more short recordings in context. The notes do not name audio formats or maximum recording lengths. A separate feature flag, features.image_generation_animation, enables animated GIF creation or editing, with optional controls for frame count and timing.
The release contains several terminal and platform changes as well. Fullscreen Codex can show image previews and pets in tmux on Ghostty and Kitty when allow-passthrough is set to all. Users can select tui.notification_method = "osc777"; approval notifications now expose longer command previews and distinguish terminal-input requests from network-access requests. On Windows, OpenAI lists fixes that stop idle local MCP servers from blocking unrelated filesystem I/O, expand deny-read globs without invoking ripgrep, and resolve filesystem permissions for sandbox policies. It also lists creation of parent directories on deny paths by the Windows MXC launcher. The announcement does not describe a new price, access tier, or general-availability change.
What the archive adds
New Runtime’s earlier coverage on November 7, 2026 described Codex CLI 0.167.0’s agent-management center, unread-result indicators, and backend switching. This is a new event in the same product’s release sequence: 0.168.0 shifts attention toward provider selection, parallel conversations, media input, and platform maintenance rather than restating those agent-management additions.
Trend signal
This is a modest signal for harness architecture outlives model choice. The provider picker puts model choice inside the Codex interaction layer, while saved per-chat settings preserve a conversation’s configuration after a global default changes. One release does not prove that the harness is the more durable boundary, but it shows the product treating provider configuration as a managed workflow concern rather than a fixed background detail.
Why it matters
Teams that route different work through different providers can now set a default for fresh Codex conversations without changing the provider recorded by an ongoing task. That separation can prevent accidental drift when a team is testing a new provider but still needs continuity on an existing debugging or implementation thread. The release notes do not claim that provider switching improves quality, latency, or cost; those remain choices for each team to evaluate.
The side-conversation workspace could help users keep a secondary investigation separate from the main task, but OpenAI provides no usage results or productivity data. Audio attachments and GIF generation expand the kinds of inputs and artifacts available inside the terminal client, yet neither feature is evidence of a broad improvement in coding-agent performance. The Windows sandbox changes are consequential for affected installations because they concern filesystem permissions and MCP-related I/O, but their real-world reliability impact is not measured in the announcement.
What happens next
- More Codex users will test provider-specific workflows in the terminal UI (3 months, 65%). Watch for: additional provider configuration or session-management controls in release notes.
- Workspace features will receive more interaction refinements (6 months, 55%). Watch for: changes to side-pane navigation, persistence, or task handoff behavior.
Limits and uncertainty
OpenAI’s release notes describe functionality and fixes, not adoption, benchmark, reliability, or productivity results. They do not identify the providers available in the new picker, supported recording formats, recording-length limits, or which users can access animated GIF generation. The workspace and animation capabilities are explicitly feature-gated, so their default availability is not established by the announcement. The Windows items identify implementation changes but do not quantify affected installations, prior failure rates, or the impact after upgrading. These operational outcomes have not been independently checked.
Skill to build now
If your team uses Codex with multiple providers, create a new chat after changing the default, then resume an existing chat and confirm the provider selected in each case before assigning consequential work. Treat that as a lightweight release regression test. Windows users relying on sandboxing should also exercise an MCP workflow and representative deny-read paths after upgrading, especially where filesystem permissions are part of the task boundary.
Reader takeaway
Codex CLI 0.168.0 makes provider choice and parallel terminal conversations more visible while adding media features and Windows maintenance fixes. The operational detail worth remembering is simple: a new-chat default does not replace the provider saved on an existing or resumed conversation. Test that boundary before relying on it in a multi-provider workflow.