Three well-named tools with clear, detailed descriptions and complete input schemas. lookup_customer and search_tickets are read-only with good parameter documentation. issue_refund is properly marked IRREVERSIBLE and includes audit trail context. However, output schemas are not documented in the visible code, only input schemas are shown. Parameter descriptions are strong (50-150 chars each), but the server lacks explicit error handling guidance and recovery patterns for common failures (e.g., customer not found, refund rejection).

• Document output schemas for all three tools. Show the structure of CustomerProfile (returned by lookup_customer), TicketResult[] (returned by search_tickets), and RefundResult (returned by issue_refund). Include field types and descriptions.
• Add error handling guidance to each tool description. E.g., lookup_customer: 'Returns 404 if customer_id does not exist. Try search_customers() with a partial name if you only have an email.' issue_refund: 'Returns 403 if customer tier does not permit refunds. Escalate to a human agent.'
• Implement a confirm_refund tool or add a dry_run parameter to issue_refund. Let agents preview the refund before committing. This prevents catastrophic errors from reasoning failures.
• Add min/max bounds to search_tickets limit parameter. E.g., 'Maximum number of results to return (1 - 100, default 5).' This prevents resource exhaustion.
• Explicitly document the audit trail in issue_refund description. E.g., 'Refund is recorded with the conversation_id from the request _meta, linking the payment back to the support interaction for compliance.' This clarifies the security model.