Scoring was not performed

• Rewrite tool descriptions in English, concisely (50-200 chars). For crawlWeb: 'Fetch and parse the text content of a single web page. Returns page title, plain text, and metadata. Fails on timeout (>30s), blocked domains, or non-HTML content.' For getIntent: 'Retrieve medical knowledge base results matching a health question. Reconstructs follow-up questions to form complete queries. Returns ranked results with source references and confidence scores.'
• Document output schemas explicitly. For crawlWeb, define: {type: object, properties: {title: string, text: string, metadata: {url, status_code, fetch_time_ms}}}. For getIntent, define: {type: object, properties: {results: [{text: string, source: string, confidence: number, doc_id: string}], query_used: string}}.
• Rename tools to verb_noun pattern. crawlWeb → fetch_webpage, getIntent → retrieve_medical_kb or classify_health_question (depends on primary purpose). Make names action-oriented and unambiguous.
• Add parameter constraints. crawlWeb's url: add format='uri', pattern='^https?://', max_length=2048. getIntent's clarified_question: add type='string', max_length=500, description='If user is following up, generate a complete medical question combining context and the follow-up. Use Chinese or English, match the user's language.'
• Add separate error handling documentation. Define three scenarios for each tool: (1) what error is returned, (2) is it retryable, (3) what should the LLM do next. Example for crawlWeb: 'Error: timeout after 30s waiting for response. Retryable: yes. Action: Try a simpler query or check if URL is valid.'