Single tool 'get_weather' has a basic description but critically lacks input schema documentation in the visible code. The description is adequate (67 chars) but the parameter 'location' has no type annotation visible in the schema structure. The tool performs a complex multi-step operation (fetch weather + air quality + RAG retrieval + LLM generation) that violates single-responsibility principle. No output schema is documented. Error handling is minimal, exceptions are raised but not categorized or recovery-guided. The tool accepts a free-form string parameter with no validation, format constraints, or enum values, inviting invalid inputs.

• Add explicit input schema with type='string', minLength=2, maxLength=100, and pattern validation for city names. Document expected format: 'City name (e.g., London, New York, Tokyo), 2-100 characters, letters and spaces only.'
• Document output schema as a structured object: {weather_report: string, location: string, temperature: number, humidity: number, aqi: string, timestamp: string}. Return JSON instead of plain text.
• Split get_weather into three tools: (1) fetch_weather_data(location) → returns temp, humidity, wind, etc.; (2) fetch_air_quality(latitude, longitude) → returns AQI components; (3) generate_weather_report(weather_data, air_quality) → returns LLM-generated narrative. This enables agent composition.
• Wrap API calls in try-catch with categorized error responses: 'Location not found. Try a major city name (e.g., London, Paris). Available: [list alternatives]' for 404; 'API rate limit exceeded. Retry in 60 seconds.' for 429; 'API timeout. Retry with a smaller location query.' for timeout.
• Add parameter validation: reject locations with special characters, validate against a whitelist of major cities, or call OpenWeather's city search endpoint first to confirm existence before fetching weather.