Skip to content

Action Cards

An Action Card is the visible review object shown before a consequential action executes.

Purpose

Action Cards answer:

  • What does the agent want to do?
  • What target will it affect?
  • What fields or state may change?
  • What is the likely consequence?
  • How reversible is it?
  • What evidence supports the proposal?
  • What decision can the user make?

Current Shape

Action Cards are defined in @agentclutch/action-card and validated by parseActionCard.

Important fields:

  • proposed_action.label
  • proposed_action.kind
  • proposed_action.target
  • proposed_action.changed_fields
  • consequence
  • risk
  • evidence
  • user_options

User Options

Current card options include:

  • approve_once
  • edit_fields
  • take_wheel
  • block
  • create_rule
  • request_more_context

Adapters map these UI decisions into loop-level ClutchDecision objects.

Validation

Use:

import { parseActionCard } from "@agentclutch/action-card";

const card = parseActionCard(JSON.parse(raw));

Invalid JSON and schema failures should be shown to the user before rendering.

Data Hygiene

Action Cards may contain sensitive context. Redact secrets before creating or recording cards.