The server defines 4 READ_ONLY tools for arxiv search with reasonable naming but significant gaps in parameter descriptions and output documentation. Tool names follow verb_noun pattern (get*) which is good. However, parameter descriptions are minimal ('The name of the papers author', 'The maximum number of results to return') and lack detail on valid ranges, formats, or constraints. Tool descriptions are present but generic (10-60 chars, below the 50-200 char baseline for LLM-optimized descriptions). Output schemas are not documented in the code, responses are inferred from implementation (formatPapers returns formatted text, not structured data). No pagination, no error classification, no guidance on retryability.

• Expand tool descriptions to 60-150 chars, explicitly stating the search dimension and use case. Example: 'Search ArXiv papers by author name. Use when you know the author but not the exact title. Returns up to 20 papers sorted by date. For broad queries, author names are case-insensitive.'
• Add parameter constraints to inputSchema and descriptions: maxResults should enforce min=1, max=100 with description '(1-100, default 10)'. Hard-cap server-side to prevent abuse.
• Fix copy-paste error in getArxivPapersBykeywords: change description from 'The name of the papers author' to 'Space-separated keywords (e.g., "neural network optimization"). Returns papers matching all keywords.'
• Document output schema explicitly. Show the structure returned by formatPapers(). Example: 'Returns text but should return JSON array: [{title: string, authors: string[], publishedDate: string (ISO 8601), arxivId: string, summary: string}]'
• Add pagination: introduce 'page' (default 1) and 'limit' (default 10, max 100) parameters. Return {papers: [...], total: number, page: number, hasNext: boolean}.
• Enhance error messages with recovery guidance. Instead of 'No papers by X', return 'No papers by X. Try: search by title/keywords, check spelling, broaden keywords. Available categories: cs.AI, cs.CV, math.AP, ...'
• Add tool annotations to all tools: readOnlyHint=true and idempotentHint=true so agents understand retry behavior.
• Add 'category' parameter enum to getArxivPapersByCategory with valid values (cs.AI, cs.CV, math.AP, etc.) instead of free-form string.

Inferred effective spec: <=2025-11-25.

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