@auctra/a2a

A2A adapter — agent-to-agent calls become Auctra Actions for the same Authority Gate.

Law: no second security model. executeAgentCall → action.execute with action_type: agent_call. Peer identity and task attenuation are enforced by the gate (Atlas delegation/a2a-peer-spoof, delegation/a2a-task-launder).

Usage

import { executeAgentCall } from "@auctra/a2a";

const decision = await executeAgentCall(auctraClient, {
  fromAgentId: caller.id,
  toAgentId: peer.id,
  claimedIntentId: intent.id,
  intentAnchorToken: intent.anchor_token,
  // Optional: downstream work the peer should run — must ⊆ caller's grant scope
  claimedActionTypes: ["create_support_ticket"],
  intent: "Hand off ticket triage",
  riskLevel: "medium",
});

What the gate checks

| Claim | Block when | | --- | --- | | Peer id | Missing, cross-org, inactive, or equals caller | | claimed_action_types | Not ⊆ caller's delegation scope.action_types (wildcards via actionTypeMatchesPattern) |

Undeclared peers stay fail-closed on agent_call (peer required). Empty claimed_action_types skips attenuation (caller grant still must cover agent_call itself).