16 math tools with complete schemas and basic descriptions. All tools have input schemas with proper type definitions (array of numbers). Descriptions are present but generic and lack LLM-optimization guidance. No parameter-level descriptions beyond the array itself. No output schema documentation. No error handling guidance. Tools follow a simple verb_noun pattern (sum, divide, sqrt) but lack context on when to use each vs alternatives. No tool annotations (readOnlyHint, etc.). Composition is sound, each tool does one thing, but descriptions don't explain dependencies or prerequisites.

• Add output schema documentation to each Tool definition. Include a 'returns' field or update tool descriptions to state: 'Returns a MathResult object with fields: result (string), operation (string), numbers (array of numbers).'
• Expand descriptions to 80-150 chars with selection context. Example: 'Sum a list of numbers. Use this for totals; use product() for multiplication. Returns the numeric sum.' This helps LLMs distinguish sum from product.
• Add minItems and maxItems constraints to the 'numbers' array schema for binary operations. E.g., divide should have minItems: 2, maxItems: 2. This prevents LLMs from passing wrong-sized arrays.
• Document error conditions in descriptions. E.g., divide: 'Divide two numbers. Returns error if divisor is zero or array length != 2.' This guides LLM error recovery.
• Add tool annotations to the Tool definition. Set readOnlyHint: true and idempotentHint: true for all 16 tools. This signals to agents that these tools are safe to retry and do not modify state.