A Arena

Reference

Hosted WebMCP tools.

Arena keeps preparation, human approval, candidate invocation, and proof verification separate so an agent cannot silently approve its own consequential action.

start_generated_release_audit

Prepares an owned Checkout release. Returns a compact record with absolute review and evidence links. It never approves or executes the candidate.

input: { version?: "vulnerable" | "fixed", idempotencyKey?: string }
nextAction: "review_and_approve_in_visible_interface"

get_generated_release_audit_status

Read-only polling. Responses include nextAction, retryAfterMs, commitments, and absolute URLs. Full evidence stays behind the linked endpoint to keep tool results compact.

Dynamically registered candidate

After visible approval, Arena registers the exact reviewed tool definition. The callback holds a single-use invocation lease, submits the reviewed arguments, and returns only after authoritative effects settle and the proof is signed.

await document.modelContext.registerTool({
  name: "preview_checkout",
  inputSchema: { /* reviewed schema */ },
  execute: ({ cartId }, { signal }) => invokeReviewedTool(cartId, signal)
}, { signal: lifecycle.signal });

Lifecycle

  1. awaiting_approval
  2. awaiting_webmcp_invocation
  3. waiting_for_effects
  4. completed or failed