Single tool 'reason' has a reasonable description (194 chars, within baseline 34-392 range) but critical schema and parameter issues. Input schema declares a 'query' object with nested 'context' and 'question' properties, but the description of 'query' itself is verbose and unclear ('Contains the following keys: context (str)...'). The nested properties lack individual type clarity in the schema presentation. No output schema is documented, LLMs cannot predict the return structure. Error handling returns unstructured text wrapped in custom XML tags rather than structured error objects with recovery guidance. The tool lacks idempotency hints, permission declarations, and rate-limit guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present.

• Document the output schema explicitly. Specify that the tool returns a string containing '<ant_thinking>...</ant_thinking>' wrapped reasoning, followed by a summary line. Include example output structure.
• Refactor the input schema: replace the verbose 'query' object description with clear nested property descriptions. Ensure 'context' and 'question' are explicitly typed as 'string' with individual descriptions in the schema.
• Replace unstructured error handling with structured error objects. Return JSON with 'error' (string), 'error_type' (enum: 'api_error', 'invalid_input', 'timeout'), and 'recovery_hint' (string). Example: '{"error": "DeepSeek API timeout", "error_type": "api_error", "recovery_hint": "Retry with a shorter query or check API status."}'
• Move DEEPSEEK_API_KEY to environment variable injection. Use `os.getenv('DEEPSEEK_API_KEY')` with a clear error message if missing, never hardcode or default to a placeholder.
• Add tool annotations: set readOnlyHint=true (tool only reads from external API), idempotentHint=true (same input always produces same reasoning output), destructiveHint=false. This guides agent planning and retry logic.