POLICIES.md

Policy Support in Kaizen

Kaizen supports storing and managing structured policies as entities. This allows agents to retrieve behavioral rules, playbooks, and guards using semantic search.

Policy Schema

Policies are stored as Kaizen entities with type="policy". The core schema is defined in kaizen/schema/policy.py.

Policy Types

Kaizen supports the following CUGA-compatible policy types:

• intent_guard: Blocks forbidden intents and returns specific responses.
• playbook: Provides step-by-step Standard Operating Procedures (SOPs).
• tool_guide: Enriches tool descriptions with additional context.
• tool_approval: Requires human confirmation before executing sensitive tools.
• output_formatter: Reformats agent responses.

Trigger Types

• keyword: Matches specific words or phrases.
• natural_language: Matches semantic meaning using vector similarity.
• always: Applies to every input/response (e.g., for global output formatters).

Using MCP Tools for Policies

The Kaizen MCP server provides several tools to manage and retrieve policies.

1. Creating a Policy

Use create_entity with entity_type="policy". The content should be the serialized JSON representation of the Policy model.

Example:

{
  "name": "Infrastructure Protection",
  "type": "intent_guard",
  "description": "Prevents unauthorized deletion of core infrastructure.",
  "triggers": [
    {
      "type": "natural_language",
      "value": ["delete the database", "wipe out the core system"],
      "target": "intent"
    }
  ],
  "content": "Access Denied: You are not authorized to perform infrastructure deletion.",
  "enabled": true
}

2. Retrieving Policies

Use get_entities with entity_type="policy" to fetch relevant policies for a given task.

# Example call
get_entities(task="I need to clear the production database", entity_type="policy")

3. Backward Compatibility

The get_guidelines tool is maintained for backward compatibility and specifically targets entities of type guideline. For policies, use the generic get_entities tool.