name:	commerce-agents-imperial
description:	IMPERIAL bounded capability profile derived from Anthropic commerce-agents. Use for shopping assistance, merchant analytics, staged catalog/inventory/pricing/campaign drafts, with human approval for writes and no autonomous payment execution.

Anthropic Commerce Agents — IMPERIAL adapter

Upstream: anthropics/commerce-agents Pinned commit: fd4d59224ab96b43c6dc6888207c67b3bd5a24cf License: Apache-2.0 Native demo prerequisites: Python 3.11+, Node 22, ANTHROPIC_API_KEY for live Claude calls.

Upstream safety model retained

The reference explicitly does not place orders, charge cards or directly change live listings. Checkout is a handoff to the host and merchant writes are staged until a person approves them.

IMPERIAL use

Install fleet-wide as TASK_GATED_COMMERCE_CAPABILITY. Only agents whose mission actually involves commerce/shopping/merchant operations should invoke it.

Allowed bounded work:

product search/comparison/planning from authorized data;
cart draft and checkout handoff preparation;
order/policy questions from verified backends;
merchant analytics and explanations;
staged inventory/listing/pricing/campaign proposals;
commerce-agent architecture, tests, evals and backend adapters.

Gates:

no autonomous card charge, checkout completion, purchase order signature or funds transfer;
no live merchant write without authenticated account authority and required human/Approval Gateway approval;
no fabricated catalog, order, customer, revenue or inventory facts;
no use of a paid Claude API under FREE_ONLY unless separately authorized by the Architect;
customer/merchant identity and authorization are host responsibilities and must be verified before protected data/actions;
credentials remain server-side and outside prompts/evidence.

Truth boundary: SKILL_BOUND != CLAUDE_COMMERCE_RUNTIME_LIVE; STAGED_WRITE != LIVE_CHANGE; CHECKOUT_HANDOFF != PURCHASE; PROPOSAL != REVENUE.