The server defines 5 tools with complete JSON schemas and descriptions. Naming is action-verb based (search_, browse_, get_), which is good. Descriptions are present and moderately detailed (90-180 chars). Schemas show proper type definitions and enums. However, output schemas are undocumented, the rubric requires explicit documentation of what each tool returns, and the source shows only input schemas. Parameter descriptions are present but generic in places ('Search query for finding...'). No error handling guidance is documented. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Limited actionable error recovery patterns. Overall: solid foundation but missing output documentation and error guidance expected of production-grade tools.

• Document output schema for each tool. Create a TypeScript interface (or inline JSON Schema example) showing the exact fields returned. E.g., search_design_knowledge returns [{id, title, category, tags, snippet, url}, ...]. Include field types and descriptions.
• Add tool annotations to the tool registration. Mark all five tools with readOnlyHint=true to signal to clients that they are safe to call without side effects and can be cached. Example: tools[0].readOnlyHint = true.
• Enhance error handling guidance in tool descriptions. Add a sentence like: 'If no results found, try a simpler or shorter query. If search times out, try filtering by category first.' This guides the LLM's retry strategy.
• Improve parameter descriptions with examples and constraints. Replace 'Search query for finding relevant design system knowledge' with 'Search query, e.g. "accessibility", "component naming", "design tokens". Supports phrase queries.'
• For search tools, explicitly document result limits and pagination. Add: 'Returns up to {{limit}} results (default 15). Use offset/page to retrieve additional results if available.' Then implement pagination in the handler.
• Add a 'notes' or 'description' field to search results to provide context snippets. This helps the LLM understand relevance before deciding whether to drill deeper. Reduces hallucination from incomplete context.
• Consider adding a 'confidence' or 'relevance_score' field to search results so the LLM can rank results by quality and filter out low-confidence matches.

Inferred effective spec: 2026-07-28+.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained