Server exposes 8 tools with complete JSON schemas and descriptions, but quality is uneven. Tools 1, 4, 5, 8 share a massive 32-parameter schema (predict_promise_miss, explain_promise_miss, get_order_risk, recommend_policy_action) with many optional fields lacking clear constraints or enums. Descriptions are present but generic (50-100 chars), offering minimal guidance on WHEN to use each tool or HOW they differ. Tool 2 (get_model_status) and 3 (get_model_metrics) are minimal stubs with empty schemas. Tool 6 (list_available_actions) and 7 (calculate_action_value) are better-scoped but still lack output schema documentation. No error handling guidance, no recovery hints, no parameter validation rules stated in descriptions. Naming is verb-first (good), but parameter names like 'seller_late_rate_7d' lack units/format hints in descriptions.

• Split the 4-tool prediction cluster into distinct tools with clear names: predict_promise_miss (returns risk score only), explain_promise_miss (returns feature importance), get_order_risk (alias for predict?), recommend_policy_action (returns action + rationale). Document in each description WHEN to call it vs the others.
• Add enums and ranges to all optional parameters. E.g., payment_type_primary: enum ['credit_card', 'debit_card', 'boleto', 'pix', 'unknown']; seller_late_rate_*: number 0-1 (fraction); customer_state: enum of valid Brazilian states or 'unknown'.
• Document output schemas for all tools. E.g., predict_promise_miss returns {risk_score: number 0-1, model_version: string, timestamp: string}. This enables LLM to chain tools and extract needed fields.
• Add error handling guidance to descriptions. E.g., 'Returns 400 if order_id is invalid or missing required fields. Try with fewer optional parameters if prediction times out. Returns 503 if model service is unavailable, retry after 30s.'
• Expand get_model_status and get_model_metrics descriptions to 100+ chars explaining what readiness/metrics mean and when to call them. Document their output schemas (e.g., readiness: {status: 'ready'|'loading', model_version: string, last_updated: string}).