Two tools with clear verb-noun naming (get_order_status, cancel_order) and explicit risk annotations. Descriptions are present and contextually rich (Italian, mentioning idempotency and approval requirements), but lack LLM-optimized English structure. Input schemas are well-formed with regex patterns for order_id validation. However, output schemas are not documented in the visible code, and parameter descriptions are minimal (only 'Order identifier matching pattern ORD-\d+'). Error handling and recovery guidance are absent from tool definitions. The server demonstrates security awareness (READ_ONLY vs APPROVAL_REQUIRED risk tags) but does not expose this via standard MCP tool annotations.

• Translate descriptions to English and optimize for LLM tool selection: 'Get the current status of a support order (read-only, no side effects)' for get_order_status; 'Cancel a support order (irreversible, requires approval)' for cancel_order. Keep to 50 - 150 chars.
• Document output schemas for both tools. E.g., get_order_status returns {status: string, created_at: ISO8601, updated_at: ISO8601, ...}. cancel_order returns {success: boolean, previous_status: string, cancelled_at: ISO8601}.
• Expand parameter descriptions: 'The order ID to query (format: ORD-NNNNNN, e.g. ORD-123456). Required.' This tells the LLM what the parameter controls and why it matters.
• Add error handling guidance to tool descriptions. E.g., 'If the order is not found, returns error code NOT_FOUND. Try searching orders by customer ID first.' Per pattern:recovery-guide.
• Declare tool annotations in MCP schema: add readOnlyHint=true to get_order_status; add destructiveHint=true and idempotentHint=true to cancel_order. This enables clients to enforce approval workflows and retry policies.