Scoring was not performed

• Expand all tool descriptions to 100-200 characters. Include WHEN to use (e.g. 'Call this when the user asks who they are or needs account details'), WHAT returns (e.g. 'Returns the authenticated Box user's name, email, and ID'), and any prerequisites or dependencies.
• Add complete parameter schemas with types, descriptions, constraints, and examples. For folder_id, specify format: 'A numeric Box folder ID as a string. Use "0" for the root folder. Example: "12345"'. For query, specify: 'Free-text search keywords. Box will match against file names, descriptions, and content. Returns up to 100 results sorted by relevance.'
• Document output schemas for each tool. E.g., who_am_i returns {user_id, user_name, user_email, login_email, status}, search_files_keyword returns {results: [{file_id, file_name, path, size, modified_date}], total_count, limit}. This enables LLMs to plan downstream calls.
• Add pagination parameters to list_folder_content_by_folder_id: limit (default 20, max 100) and offset (default 0). Document expected return structure including total_count so LLMs can fetch all items if needed.
• Implement structured error responses from mcpClient.callTool(). Instead of catch-and-log, return {error: 'Tool failed', code: 'AUTHENTICATION_REQUIRED|NOT_FOUND|INVALID_INPUT', message: 'User not authenticated with Box. Please re-login.', recovery: 'Try calling who_am_i to verify authentication status.'}. Pass this structured error back to OpenAI so the LLM can take corrective action.
• Add the missing 'question' parameter to ai_qa_single_file. Currently only file_id is documented but the tool name and system prompt strongly suggest users must ask a question about the file. Schema should be: {file_id: {type: 'string'}, question: {type: 'string', description: 'Your question about the file content (e.g., "What is the main topic?" or "Summarize the key points")'}}.