Server has 7 tools with complete input schemas and descriptions. Naming follows verb_noun pattern (decide_, search_, get_, report_). Descriptions are substantive (100-200 chars) and explain WHEN to use each tool. All parameters have types and descriptions. However, output schemas are not documented in the source code, only input schemas are visible. Error handling is present (gate denials, quota checks) but recovery guidance is minimal. Tool annotations (readOnlyHint) are correctly applied. The server is well-structured for Korean public contract domain but lacks explicit output schema documentation and some parameter constraints could be tighter.

• Document output schemas for all tools. For search_* tools, specify: {results: [{id, title, excerpt, source, relevance_score}], total_count, next_cursor}. For get_* tools, specify the full article/case structure returned.
• Add conditional parameter documentation to decide_contract_method: 'service_type is required only when contract_type=service; construction_specialty is required only when contract_type=construction.'
• Define explicit result limits in search_* tool descriptions: 'Returns up to top_k results (max 20). If more results exist, use next_cursor to paginate.' Include next_cursor in output schema.
• Enhance error responses with actionable recovery hints. E.g., 'attribution_error: Could not identify caller. If using stdio, this is expected. If using HTTP, check request headers.' instead of silent failures.
• Replace report_issue context parameter type 'object' with a defined schema: {tool_name, arguments, error_message, timestamp} so LLMs know what to include.