Four tools with complete input schemas and detailed descriptions. All tools follow verb_noun naming (search_paper, get_paper, get_authors, get_citation). Descriptions are substantive (100-200 chars), exceeding the 10-1024 char baseline. Parameters are well-typed with enums (format in get_citation). However, output schemas are not documented, LLMs cannot predict response structure for downstream chaining. No error handling guidance. Missing tool annotations (readOnlyHint, idempotentHint). All tools are read-only but this is not declared in metadata.

• Document output schemas for all four tools. Example: 'search_paper returns {results: [{paperId, title, authors, year, citationCount}], total: number, offset: number}'.
• Add tool annotations: readOnlyHint=true and idempotentHint=true to all four tools to signal safety and retry eligibility.
• Add error handling guidance to tool descriptions: 'If paper_id is invalid, returns 404 with available ID formats. If rate limited, retry after 60 seconds.'
• Clarify search_paper pagination: 'Results are limited to 100 per request. Use offset to paginate. Returns total count for planning.'
• Document which 'fields' are expensive or require API keys (e.g., embeddings). Guide agents to request only necessary fields.