Ponytail Codex Setup with the Native Plugin

Use a Codex version that exposes the plugin and marketplace commands in its official documentation. Confirm which installation and user profile the command reaches, especially if you have several command-line or Desktop environments. Record your operating system, Codex version, and the plugin's source before changing configuration.

The Ponytail reference used here is 5.3.2 at commit 09c434a. Read the upstream Codex installation section and review the hook scripts that will run locally. A skill is an instruction resource; a hook can execute code at lifecycle events, so hook trust is a separate decision.

Preserve existing AGENTS.md files and plugin configuration. Start with a disposable repository and a recoverable configuration rather than using private production code to test whether the installation loads.

Run these commands in a system terminal:

codex plugin marketplace add DietrichGebert/ponytail
codex plugin add ponytail@ponytail

Read the output for the registered source and any reported errors. These are Codex CLI commands, not the Claude Code in-chat installation syntax. Follow the upstream's restart instructions for Codex Desktop after installation.

Inside Codex, open /hooks, inspect the two lifecycle hooks from the plugin, and review what each executes. Trust them only when the displayed source and behavior match the code you reviewed. Then begin a new conversation. Adding a plugin registration without the required hook trust or a fresh lifecycle is not the same as completing activation.

The command syntax is also described in the OpenAI plugin reference and marketplace reference. If your installed host does not expose these commands, resolve the host-version difference rather than inventing an alternative command.

On macOS with Codex CLI 0.137.0, an isolated configuration added the marketplace and plugin, listed the installed 5.3.2 plugin, refreshed the marketplace, and removed both registrations. The installed plugin's tracked source files matched our fixed reference at the time of the check.

A pinned marketplace does not necessarily pin the plugin: the checked marketplace manifest points its plugin source at main. Inspect the actual installed source and version rather than assuming a marketplace ref makes every future installation immutable.

We did not authenticate an agent session, confirm hook trust, execute the lifecycle hooks, or verify namespaced skill loading in a new conversation. These installation-management checks are not an end-to-end activation pass.

After restart or reload, check that the plugin is registered and the reviewed hooks are trusted. Start a new chat and inspect the skills available from the plugin. Codex uses a plugin namespace; the upstream gives this review example:

$ponytail:ponytail-review

Do not assume /ponytail-review from a Claude Code example is the equivalent Codex invocation. Use skill discovery in the host and the upstream command mapping for the installed version.

For a behavior check, ask for a small change that can reuse an existing project helper. Specify the acceptance conditions, public interface, edge cases, and relevant tests. Inspect the plan and diff, then run the checks. An agent saying that it is following a skill is weaker evidence than the host showing the resource is loaded, and neither replaces code review.

Keep the actual environment and observations in your record. A successful local CLI check does not establish that a different Desktop process or user profile loaded the same plugin.

Codex can also read repository instructions from AGENTS.md. Ponytail supplies a compact upstream rules file. Review and merge its guidance into the appropriate project instructions instead of replacing a file that already contains team requirements.

Instructions are affected by directory scope and other applicable project files. Read the OpenAI AGENTS.md guide when deciding where a rule belongs. Place project-specific constraints where they apply, and keep instructions to preserve necessary validation, compatibility, tests, and accessibility.

This fallback provides rules. It does not create plugin hooks, plugin skill discovery, or a complete command suite. It also does not gain a reliable off switch from the plugin's mode controls. Remove or revise the specific instruction text to change a rules-only setup, then start a fresh conversation and inspect the applicable rules.

The upstream's default level is full. Shared user configuration can select lite, full, ultra, or off for new hook-initialized sessions. The environment variable PONYTAIL_DEFAULT_MODE takes precedence over the shared configuration file's defaultMode.

Use the defaults and platform paths to set that preference deliberately. A change in one conversation should not be assumed to persist to every new session. A static AGENTS.md fallback continues to supply its text independently of the hook configuration.

The documented marketplace refresh is:

codex plugin marketplace upgrade ponytail

This refreshes the Git marketplace. It does not by itself prove that the running plugin has been replaced. A deliberate refresh-and-reinstall workflow combines the documented refresh with removal and addition:

codex plugin remove ponytail@ponytail
codex plugin add ponytail@ponytail

This is a sequence of supported operations, not a claim that Codex has a dedicated plugin update command. Review hook trust again and start a new chat after reinstalling.

To uninstall the plugin without adding it again, use codex plugin remove ponytail@ponytail. Restore only instructions you edited in project files. The upstream's separate full cleanup script can affect shared defaults, mode markers, and other local integrations; inspect its scope before using it as part of a complete cleanup. There is no universal downgrade command established by these instructions.

If hooks are missing, a skill name is unavailable, or behavior differs, check host version, source, trust, profile, and session age before assuming a problem with the coding rules. Continue to the troubleshooting guide for instruction conflicts and task-level risks.