Add Rule Command

You are helping the user add command permission rules to both Claude Code and Codex.

How rules work:

• Claude Code: Uses permissions in settings.json with allow, ask, deny arrays
• Codex: Uses Starlark rules in ~/.codex/rules/*.rules with prefix_rule() function

Both systems control which shell commands can run and whether to prompt for confirmation.

Arguments Parsing

Parse $ARGUMENTS for rule details:

• Command pattern: Which command(s) should this rule match?
• Decision: allow, prompt/ask, or deny/forbidden?
• Reason/justification: Why is this rule needed?

Flags

Check for these flags in $ARGUMENTS:

• --claude-only: Only configure Claude Code
• --codex-only: Only configure Codex
• --project: Use project scope for Claude (Codex rules are always user-scope)

Decision Mapping

User Intent	Claude (permissions)	Codex (Starlark)
"allow", "permit", "always run"	allow array	decision = "allow"
"ask", "confirm", "prompt"	ask array	decision = "prompt"
"deny", "block", "forbid", "never"	deny array	decision = "forbidden"

Flow

Step 1: Gather Rule Intent

If the user provided sufficient context, proceed to Step 2.

If missing required information, ask conversationally:

"I'll help you add a command permission rule. I need to know:

• What command should this rule match? (e.g., 'git push', 'rm -rf', 'npm publish')
• What should happen? Allow it, ask for confirmation, or block it?
• Why? (optional but helps document the rule)"

Step 2: Generate Claude Permission

Target file:

• User scope: ~/.claude/settings.json
• Project scope: .claude/settings.json

Format:

{
  "permissions": {
    "allow": ["Bash(npm test:*)"],
    "ask": ["Bash(git push:*)"],
    "deny": ["Bash(rm -rf:*)"]
  }
}

Pattern syntax:

• Bash(command:*) - Match command with any arguments
• Bash(git push:*) - Match "git push" with any args
• Bash(npm run:*) - Match "npm run" with any args
• Bash(rm -rf:*) - Match dangerous rm commands

Step 3: Generate Codex Starlark Rule

Target file: ~/.codex/rules/meta-unify.rules

Format:

# Rule: DESCRIPTION
prefix_rule(
    pattern = ["command", "subcommand"],
    decision = "prompt",
    justification = "REASON_HERE",
)

Pattern syntax:

• Pattern is a list of command elements: ["git", "push"] matches git push
• Can use unions for alternatives: ["git", ["push", "force-push"]]
• More specific patterns take precedence

Decision values:

• "allow" - Run without prompting
• "prompt" - Ask before each invocation
• "forbidden" - Block without prompting

Step 4: Apply Changes

For Claude:

1. Read existing settings.json
2. Merge into appropriate permissions array (allow/ask/deny)
3. Validate JSON syntax
4. Write updated config

For Codex:

1. Check if ~/.codex/rules/ exists, create if not
2. Read or create meta-unify.rules
3. Append new rule (avoid duplicates)
4. Validate Starlark syntax
5. Write updated file

Step 5: Handle Partial Failure

If one system fails:

"Claude rule added successfully, but Codex rule failed: [ERROR] Keep Claude changes? [y/n]"

Step 6: Confirm

Report success:

"Added permission rule for 'git push':

Claude Code: Added to ask array in ~/.claude/settings.json

"ask": ["Bash(git push:*)"]

Codex: Added to ~/.codex/rules/meta-unify.rules

prefix_rule(
    pattern = ["git", "push"],
    decision = "prompt",
    justification = "Confirm before pushing to remote"
)

Both systems will now ask for confirmation before running git push."

Examples

User input: "Require confirmation before git push"

• Command: git push
• Decision: prompt/ask
• Claude: "ask": ["Bash(git push:*)"]
• Codex: prefix_rule(pattern=["git", "push"], decision="prompt")

User input: "Block all rm -rf commands"

• Command: rm -rf
• Decision: deny/forbidden
• Claude: "deny": ["Bash(rm -rf:*)"]
• Codex: prefix_rule(pattern=["rm", "-rf"], decision="forbidden")

User input: "Always allow npm test"

• Command: npm test
• Decision: allow
• Claude: "allow": ["Bash(npm test:*)"]
• Codex: prefix_rule(pattern=["npm", "test"], decision="allow")

Rule Precedence

• Claude: deny > ask > allow (most restrictive wins)
• Codex: forbidden > prompt > allow (most restrictive wins)

Reference

For complete format specifications, invoke the meta-unify-core skill to access:

• references/claude-formats.md (permissions section)
• references/codex-formats.md (rules section)