Strong schema definitions with proper JSON Schema validation, clear parameter constraints (minLength, maxLength, enums), and comprehensive input validation. Tool names follow verb_noun convention (get_, search_, evaluate_, find_, apply_, prepare_, request_, revise_). Descriptions are present and actionable (avg ~120 chars). However, output schemas are not documented, LLMs cannot predict response structure. Parameter descriptions are minimal or absent for many fields (e.g., 'lines' array lacks explanation of what each line represents). Error handling guidance is not visible in tool definitions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications (READ_ONLY vs WRITE). Missing idempotency guidance despite idempotencyKey parameters in 3 tools.

• Document output schemas for all 8 tools. Example: 'Returns {decisionId: string, violations: [{code: string, severity: string, message: string}], repairOptions: [{repairId: string, description: string}]}'. This enables LLMs to chain tools correctly.
• Add descriptions to all nested object properties. For 'lines' array: 'Array of line items, each with productId (SKU or catalog ID) and quantity (units to purchase). Required; at least 1 item, max 20.' For 'budget': 'Shopper budget ceiling with amount (0 - 100000) and currency (ISO 4217 code, e.g. USD, EUR).'
• Declare tool annotations in the MCP server registration. Mark apply_plan_repair, prepare_validated_cart, request_quote, revise_validated_cart with destructiveHint=true. Mark apply_plan_repair, prepare_validated_cart, revise_validated_cart with idempotentHint=true (since they accept idempotencyKey).
• Add error handling guidance to tool descriptions. Example for apply_plan_repair: 'Returns 400 if decisionId is invalid or expired (retry with fresh evaluate_shopping_plan). Returns 409 if repairId no longer applies (call find_valid_alternatives for current options). Returns 422 if lines do not match pre-repair state (idempotencyKey prevents duplicate application).'
• Standardize parameter descriptions across tools. Create a shared glossary: decisionId = 'Unique decision identifier returned by evaluate_shopping_plan; valid for 5 minutes', idempotencyKey = 'Caller-generated UUID or nonce (8 - 160 chars) to ensure this operation is applied at most once, even if retried'. Reference it in each tool.