This MCP server exposes 31 tools for SQL Server performance analysis. Strengths: all tools have descriptions (50-200 char range), clear naming conventions (verb_noun pattern with domain-specific prefixes), structured input schemas using Zod validation, and comprehensive artifact routing via route_artifact. Weaknesses: tool schemas lack output documentation, no explicit error recovery guidance, missing parameter descriptions in some tools, and no pagination or result limit documentation despite potential for large dataset returns. The server is well-intentioned and domain-expert (SQL Server performance tuning is highly specialized), but falls short of production-grade patterns for LLM orchestration: tools that accept arbitrary 'input' strings need stronger validation and error messages, and skill-based tools lack concrete output schema definitions to guide downstream LLM reasoning.

• For all 27 skill-review tools, document the output schema in code comments and return metadata. Example: `{ type: 'text', text: '...', metadata: { checksRun: number, failuresFound: number, outputFormat: 'markdown|json' } }`. This enables LLMs to understand what data is available.
• Add input validation with actionable error messages. Example: 'Input exceeds 1MB limit; truncate to first 1000 lines' or 'Input does not appear to be valid XML query plan, expected <ShowPlanXML> root element'.
• Implement result limiting and pagination for list_skills. Add parameters: `limit` (1 - 100, default 20) and `offset` (default 0). Return total count and next_offset for chaining.
• Enhance error messages with 'Did you mean?' suggestions. When a skill name is not found, compute edit distance and suggest closest matches from available skill names.
• For review tools that return structured analysis (checks, findings, thresholds), define a canonical output schema and include it in the tool definition. Example: `{ checks: [{ name, status, severity, explanation }], summary: { checksRun, failuresFound, severity } }`.
• Document the optional 'input' parameter behavior more explicitly: 'Omit to retrieve skill definition only; provide artifact content to run analysis. Supports up to 10MB plain text (SQL, XML, logs, CSV data).'
• Add tool annotations (readOnlyHint=true, idempotentHint=true) to all tools, as they are read-only and safe for agents to retry without side effects.
• Consider adding a 'get_skill_example' tool that returns sample input and output for a given skill, guiding LLMs on how to call skill-review tools effectively.

Inferred effective spec: 2026-07-28+.

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