codeArbiter Decision and First-Use Guide: Bringing Gated Code Governance to Claude Code, Codex CLI, and Pi

Who it's for

• Developers and teams working in Claude Code, Codex CLI, or Pi who want changes checked before they reach the codebase: codeArbiter is a governance layer designed for those three hosts, and the official repository title states its motivation — "When you can't trust yourself with your code base, trust Arbiter."
• Once enabled, every intent you express to the agent is routed to a gated skill or reviewer agent; nothing commits until all gates are green; decisions go through SMARTS; the audit trail is append-only. It operates at repository granularity: the opt-in target is a specific repository. If you need mandatory pre-commit gates and traceable decision records, this is built for exactly that need. Append-only means audit records can only be added, never deleted or rewritten.
• Suited to cautious, look-before-you-leap users: the official documentation states explicitly that the plugin stays dormant after install and does nothing to a repository until you explicitly opt that repository in — the original text reads "That silence is intentional." Installing it does not immediately change the repository's existing behavior.
• When it is not a fit: workflows that don't want a gate-based process, one-off code analysis or rewriting tasks, and machines whose environment can't meet the requirements in the next section. Which programming languages and project types it applies to is not covered by the official documentation and remains unknown, and that directly affects whether you should pick it.
• Whether agent hosts beyond Claude Code, Codex CLI, and Pi are supported is not stated in the evidence and remains unknown.

Your first attempt in 3 minutes

• Prerequisites: Windows, macOS, or Linux; the official requirements are Pi 0.80.10, Node.js 22.19 or newer, Python 3, and Git, with no account or sample data needed. If any one is missing, set the environment up first and install afterwards.
• Where to install: the officially documented install path is the Claude Code plugin marketplace. In Claude Code, run the following two commands in order — the first adds arbiterForge/codeArbiter as a marketplace source, and the second installs the plugin named ca from that source:

/plugin marketplace add arbiterForge/codeArbiter
/plugin install ca@codearbiter

• Success signal after install: hooks, commands, agents, and statusline wiring load automatically, and every command resolves under the /ca: namespace — in other words, no manual wiring is needed after install, and commands prefixed with /ca: should be recognized. At this point the plugin is still dormant; that is expected behavior, not a failed install.
• Key confirmation point: a repository enters the governance flow only after it is explicitly opted in — a finished install is not the same as a working setup. The specific opt-in command does not appear in the official documentation; how this step is performed is unknown, so defer to the instructions inside the official repository and do not guess at commands.
• Your first task: after opting in, express an ordinary code analysis or modification intent to the agent, then watch two things — whether the intent is routed to a gated skill or reviewer agent, and whether nothing commits before all gates are green. If the intent is not routed, or a commit lands while gates are not all green, the governance flow is not running as officially described. Which specific commands exist under /ca: is not covered by the evidence and cannot be quoted verbatim.

Use cases

• Case 1: Enable gated commits on a personal repository. Scenario: you want every change automatically checked before it is committed. Approach: run the two install commands above in Claude Code to install the ca plugin, confirm that hooks, commands, agents, and statusline wiring have loaded automatically and that /ca: commands resolve, then opt the repository in, and finally express a change intent to the agent. What to expect: the intent is routed to a gated skill or reviewer agent, nothing commits until all gates are green, decisions go through SMARTS, and audit records can only be appended. Because the plugin is dormant after install, every side effect in this case begins at the opt-in step; the install commands themselves trigger no governance actions. If you only want to see what the installed state looks like, you can stop at "wiring loaded automatically, /ca: commands resolve" and observe — there is no need to rush into opting in.
• Case 2: Make your team's change decisions traceable. Scenario: with several people collaborating, you want every change decision on record and its history safe from rewriting. Approach: in a repository that is already opted in, express a change intent to the agent and run the same gated workflow. What to expect: decisions are recorded through SMARTS, the audit trail is append-only, and records already written are not overwritten by later operations. Exactly how interaction with SMARTS works and where to view the records are not stated in the evidence; the confirmed scope ends there.

Limits

• Side effects and confirmation points: installing automatically loads hooks, commands, agents, and statusline wiring — officially documented behavior that comes with installation. The initial state is dormant, and until a repository is explicitly opted in the plugin does nothing to it, so without opting in there is no way to verify the governance effect. Whether an opt-in can be undone, and how, is not stated in the evidence and remains unknown; if rollback cost matters to you, settle this point before you opt in.
• Files, commands, and network: whether it reads or writes your files, which commands it executes, and whether it touches the network are all not covered by the official documentation and remain unknown; that silence is no proof that it has no such behavior. Before connecting it to an important repository, try opting in inside a disposable environment first.
• Platform and client differences: the only install path given verbatim by the official documentation is the Claude Code plugin marketplace. Codex CLI and Pi are named as hosts the governance layer serves, but their install methods do not appear in the official documentation and remain unknown.
• Key information not yet confirmed: the specific opt-in command, the full set of commands under the /ca: namespace, how interaction with SMARTS works, supported programming languages and project types, support beyond Windows, macOS, and Linux, and how to uninstall or temporarily disable it — none of these appear in the evidence. Before adopting it, confirm each point at the official repository.