How to use the Applying a permission policy agent skill

What is an agent skill and what does this one do?

An agent skill is a folder with a SKILL.md file that gives an AI agent instructions for one kind of task. This skill covers one task: configuring command permissions for an OpenRig agent team in the harness that runs the agent.

The skill explains the permission options, recommends keeping the safe team default, and only adds allow rules you explicitly approve. It backs up files before editing them and avoids adding duplicate entries. It also states that some lifecycle commands still ask for approval even after you say yes.

Install the skill in Claude Code or the Claude apps

Download the ZIP file from the skill's page on this directory and unzip it into ~/.claude/skills/ for your user account, or into a project's .claude/skills/ folder to limit it to one project.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Requirements: OpenRig installed, an agent harness such as Claude Code or Codex, and write access to the relevant native settings files (project or user config).

Trigger the skill with natural requests

You do not need a special command. The skill is meant to be used before launching a team during agent-guided setup, or when you ask the agent to configure OpenRig command permissions, reduce repeated native approval prompts, or apply a selected rig or seat permission policy.

• Configure OpenRig command permissions for this project.
• Reduce the repeated approval prompts my agent team keeps showing me.
• Apply the permission policy I chose for this seat.
• Set up the team before we launch it, and ask me which permissions I want.

Choose the scope of permissions

The agent explains the options and asks which one you want unless an explicit choice already exists. The source describes four choices: keep the team default, prompt for everything (only on explicit request), remember selected commands, or broader permissive operation.

• Keep the team default: leaves an unset policy unset and preserves your current native settings. Claude lifecycle asks remain.
• Prompt for everything (none): an explicit opt-out of the team allowances. Native rules still decide each prompt.
• Remember selected commands: adds native allow rules for the chosen command family or narrower verbs, leaving other rules intact.
• Broader permissive operation: explains filesystem and network exposure and needs its own explicit opt-in.

Understand what the agent changes and how you undo it

Before writing, the agent identifies the target settings files, reads existing rules, and prepares a diff. It preserves deny and ask rules, sandbox posture, hooks, auth, MCP and model settings. Backups are made of touched files, and only authorized additions are merged.

To reverse the changes, use this request: Undo the OpenRig command allowances added by this setup; keep my other rules. The agent removes only the recorded entries from their exact files. For Claude it removes the permissions.allow entries it added; for Codex it removes the prefix_rule entries.

• Whole-family allows match all rig verbs, including lifecycle and topology changes. Narrower prefixes such as rig ps or rig queue list are offered when they fit better.

Check that the change loaded

After applying a change, the agent validates the format and confirms whether your version needs a new session. It then verifies an ordinary matching command twice in the target conversation and checks that an unrelated command gained no matching rule, using harmless reads only.

A file write does not prove that an existing conversation loaded the change, and a command already allowed by the launch default does not prove that a new rule is active.