Three tools are registered via @mcp.tool() decorators with names, input schemas, and descriptions. Schemas are properly typed with descriptions. However, output documentation is missing entirely, the rubric requires documented output schemas (DIMENSION 1.D critical check). Descriptions are adequate but brief (avg ~100 chars). Parameter naming follows verb_noun convention (search_*, extract_*), though parameter descriptions lack range/format guidance. No error handling guidance, no idempotent/destructive hints, no confirmation patterns for state-modifying operations. Resources and prompts are present but not core tools. Overall: decent foundation but significant gaps in output documentation and error recovery patterns.

• Add documented output schemas to each tool description. E.g., 'Returns a JSON object with fields: paper_id (string), title (string), authors (array), summary (string), pdf_url (string), published (date). For search_papers, returns an array of paper IDs.' See pattern:tool.
• Update search_papers description: 'Searches arXiv and SAVES results to disk under papers/<topic>/papers_info.json. This is a state-modifying operation, results are persisted.' This tells the LLM it has side effects.
• Add range constraints to max_results and num_papers parameters. E.g., 'Maximum number of results to return (1-100, default 5)' and 'Number of papers to search for (1-50, default 5)'. See pattern:constrained-input.
• Add error handling text to tool descriptions. E.g., 'If the topic is not found on arXiv, returns an empty list. If filesystem write fails, returns error message starting with "Error:". If this occurs, check disk space and retry.' See pattern:recovery-guide.
• Declare tool annotations in the schema (via fastmcp if supported, or inline descriptions). For search_papers: 'This tool modifies server state; consider it destructive. Agents should not retry without user confirmation.' This prevents accidental duplicate calls.
• Refactor extract_info to return structured JSON with typed fields, not a free-text JSON string. Return a dict/object directly: {"paper_id": "...", "title": "...", ...} so LLMs can key on field names.
• For all tools, add 'If the resource is not found or unavailable, the error message includes available alternatives.' This guides LLM error recovery. See pattern:recovery-guide.