Two well-defined tools with clear, actionable descriptions and complete input schemas. request_approval has comprehensive parameter documentation with proper constraints (min/max for ttl_minutes and wait_minutes). Both tools follow verb_noun naming convention. Schemas use Zod with type safety and descriptions. However, output schemas are not formally documented, responses are text-only without structured field definitions. Error handling is implicit (relies on Confirm SDK exceptions) rather than explicit in tool descriptions. No tool annotations (readOnlyHint/destructiveHint) despite request_approval being irreversible.

• Add formal output schema documentation. Define that both tools return {status: 'PENDING'|'APPROVED'|'REJECTED'|'EXPIRED', id: string, text: string, effectivePayload?: unknown}. This enables agents to extract and chain results.
• Add tool annotations: request_approval should have destructiveHint=true; check_approval should have readOnlyHint=true. Use MCP SDK's annotation support to declare safety properties.
• Expand check_approval description to explain: 'Poll a pending approval request by id. Returns PENDING (still waiting), APPROVED (use effectivePayload), REJECTED, or EXPIRED. Call this if request_approval times out and you need to check later.'
• Document error cases in descriptions: 'If the approval ID is invalid or expired, returns EXPIRED status. If the Confirm service is unreachable, the tool will fail, retry after a delay.'
• Clarify idempotency: 'Calling request_approval twice with identical parameters within the same ttl_minutes window returns the same request ID. Safe to retry on transient failures.'