Maps severity levels to policy actions for the InputGuard plugin.

The policy determines what happens when input is flagged at each severity level.

Actions

• :block — Halts the agent loop by setting needs_response: false and injecting an assistant message indicating the request was blocked.
• :warn — Injects a system message warning the LLM about the flagged input. Execution continues normally.
• :log — Logs the violation via Logger.warning/1. Execution continues unchanged.
• :callback — Calls the user-provided on_violation function from config.
• fun/2 — A function fn result, ctx -> ctx end for fully custom handling.

Default Policy

%{suspicious: :warn, blocked: :block}