Critical issues: Only 2 unique tools exist (search_papers, extract_info), but the tool list shows 6 duplicates, indicating either a registration bug or inference error. Both tools have basic descriptions (60-80 chars, below the 194-char baseline), but lack parameter descriptions entirely. Input schemas are present with type info, but parameter descriptions are missing. No output schema documentation. No error handling guidance. No tool annotations. The server lacks composition patterns, search_papers writes to disk but extract_info only reads, yet they're not clearly distinguished in descriptions. No pagination support despite potentially returning multiple papers.

• Add parameter descriptions to both tools: 'topic' → 'The research topic to search for on arXiv (e.g., machine learning, quantum computing)'; 'max_results' → 'Maximum number of papers to retrieve (1-50, default 5)'; 'paper_id' → 'The arXiv paper ID (e.g., 2301.12345)'.
• Document output schemas: search_papers returns {type: array, items: {type: string, description: 'arXiv paper ID'}, example: ['2301.12345', '2302.67890']}. extract_info returns {type: object, properties: {title, authors, summary, pdf_url, published}, or error string}.
• Add error recovery guidance: extract_info should return 'Paper {paper_id} not found. Try search_papers(topic="...") to find papers first, then use extract_info with the returned paper ID.'
• Implement pagination: Add 'limit' (1-50, default 20) and 'offset' (default 0) parameters to extract_info if it searches across multiple topics. Return {papers: [...], total: N, offset: M} to enable iteration.
• Add tool annotations: Mark search_papers with destructiveHint=true (writes to disk), extract_info with readOnlyHint=true. This helps agents understand side effects.