Codex CLI 0.166.0 Adds Terra and Network Diagnostics

Codex CLI 0.166.0 was released on October 29 with GPT-6.1 Terra catalog support, keyboard transcript selection and a focused network diagnostic command. It also preserves acknowledged permission settings when resuming or forking threads and improves reconnection recovery. For terminal users, the update connects a new model option with practical controls for navigating long conversations and diagnosing connection failures.

Codex CLI 0.166.0 adds Terra to model catalogs

OpenAI’s changelog lists Terra in bundled, Vertex AI and Amazon Bedrock catalogs. Users can start a session with codex -m gpt-6.1-terra when the provider and account have access. This is a CLI catalog change following the October 26 Terra release, rather than a separate model launch. Workspace permissions and provider entitlements still determine availability.

Terra is positioned for writing, visual reasoning and computer use. To evaluate it in a repository workflow, keep the task and validation steps the same as an existing Sol run. Ask for a bounded change, inspect the diff and run the checks that establish whether the requested behavior works. A different default model can affect both the result and resource use, so a known task makes the comparison useful.

The GPT-6.1 Terra reference covers model settings and prices. The access guide explains the account and product checks. A model listed in a terminal picker is a configuration option, not confirmation that every provider has enabled it for the current account.

Navigate and select transcript text from the keyboard

The new selection mode supports word, line and page motions, integrated search and configurable shortcuts. This matters for long terminal sessions in which an engineer needs to copy a command, compare a previous answer or inspect an earlier error without leaving the keyboard. Rendered Markdown is also made consistent across asynchronous questions, choices, answers and replies.

After upgrading, try selection on a saved conversation containing prose, a command and a link. Check that the selected text contains the intended range and that line wrapping has not changed its meaning. For commands, inspect the copied text before executing it. A transcript is evidence of what happened during the session; copying from it should preserve the context needed to use that evidence correctly.

Teams with customized keybindings should inspect their effective shortcuts rather than assuming that another user’s terminal behaves identically. Keep a small note of the selection and search controls that work in the team’s supported terminal. That makes the feature useful during an incident, when navigating a large conversation is more urgent than learning a new shortcut scheme.

Diagnose network configuration with a focused command

The release adds codex doctor --network-only for configuration, connectivity and WebSocket diagnostics. Use it to narrow a connection problem before changing the project’s code or replacing credentials. Save the relevant error and identify the account, provider and environment in which it occurred; a successful connection from a browser does not establish the same path from a terminal process.

For an intermittent failure, compare a failing attempt with a successful one under the same configuration. Check whether a proxy, VPN, remote host or organization gateway changes the request path. The diagnostic command is a tool for gathering evidence. Its result should guide the next investigation step rather than prompting a broad reset of every setting.

This is particularly useful for a CLI used across local and remote environments. A repository can be available while the model service is unreachable, and a service connection can work while a separate integration fails. Keeping the failure’s actual command and environment visible helps distinguish those cases. These are troubleshooting recommendations, not a claim that the new command automatically repairs a connection.

Resume work with the intended permissions and context

OpenAI lists fixes that preserve acknowledged approval and permission settings when threads are resumed or forked. The release also preserves local image attachments when restoring killed composer text and reasoning-token accounting across resume and fork. Other fixes address descendant-thread restoration, application-server lifecycle races and TUI reconnection behavior, including offline quit and exit handling.

For a resumed task, confirm the active repository, model and permission profile before issuing a consequential command. Then inspect the previous changes and identify the next unfinished action. A preserved setting improves continuity, but the user still needs a clear understanding of the state in which work continues. For image-based tasks, check that the restored attachment is the intended source before asking the agent to use it.

When a workflow uses reusable prompts, keep the expected inputs and review checks with the prompt. The site’s skills directory offers examples of packaging repeatable work around those checks. Permissions describe what the agent may do; the task instructions and validation determine whether it did the intended work.

Upgrade a working installation and verify one task

The official release command is npm install -g @openai/codex@0.166.0. Use the installation method appropriate to the existing setup, then check codex --version and open a small known task. The version check confirms the executable being run, which is useful when more than one installation is present on the command path.

Validate the features that matter to your workflow: model selection, transcript navigation, resume behavior or network diagnostics. Keep the upgrade’s acceptance check small enough to repeat on another machine. The release supplies several improvements, but a targeted check against the team’s actual terminal and provider is the evidence needed to continue work confidently.