14 tools with consistent naming (verb_noun pattern: search_, clickbus_), good descriptions (100-200 chars typical), and explicit input schemas with enums. All tools declare risk levels and annotations. However, parameter descriptions are sparse or missing for many inputs, output schemas are undocumented, and error handling lacks recovery guidance. Privacy-mode and response_format parameters appear on most tools but lack descriptions explaining their effects. No documented return types or field mappings for chaining.

• Document output schema for every tool. For clickbus_search_trips_v5, specify: returns array of {trip_id, from, to, departure_time, available_seats, total_seats, price_brl, operator}. Include field types and which IDs are needed for downstream calls (e.g., trip_id for clickbus_book).
• Add descriptions to privacy_mode and response_format parameters. E.g., privacy_mode: 'Controls data redaction level: summary (minimal PII), structured (full data), raw (unredacted). Default: summary for privacy.' response_format: 'Output format: markdown (human-readable) or json (structured). Default: markdown.'
• Implement structured error responses with classification. Return {error_code, message, category: 'retryable'|'user_fixable'|'fatal', next_steps: ['...'], suggestion: '...'}. E.g., clickbus_book error: {category: 'user_fixable', message: 'Mutations disabled', next_steps: ['Set CLICKBUS_ALLOW_MUTATIONS=true'], suggestion: 'Ask user to enable mutations'}.
• Add pagination parameters to list/search tools. clickbus_search_trips_v5 should accept limit (1-100, default 20) and offset (default 0). Return total_count and has_more in response so agents can iterate safely.
• Implement confirmation flow for destructive operations. clickbus_book and clickbus_cancel should return a confirmation token on first call (with dry-run results), requiring that token + explicit_user_intent on second call to execute. Prevents accidental double-booking.