The 'roll' tool has a well-structured schema with clear parameter types and detailed descriptions of supported dice notation. However, the tool definition lacks critical information: no output schema documentation, no error handling guidance, and the description, while informative, could be more action-oriented for LLM selection. The server properly implements JSON Schema with type constraints and a comprehensive notation guide, but falls short of production-grade quality by not documenting what the agent should expect in responses or how to handle parsing failures gracefully.

• Document the output schema explicitly: 'Returns {expression: string (the parsed input), breakdown: string (step-by-step roll results), total: number (final computed value)}.' This allows agents to extract and chain data.
• Enhance error messages with actionable guidance: 'Invalid expression "3d6x": unknown modifier "x". Did you mean "3d6!" (explode) or "3d6e4" (explode on 4+)?' Reference the pattern 'recovery-guide'.
• Rewrite tool description in 50-150 characters for agent clarity: 'Roll dice with support for modifiers (keep/drop, explode, reroll), math operations, and complex expressions. Use for tabletop games, probability checks, and random number generation.'
• Clarify the optional 'description' parameter: 'An optional label for the roll context (e.g. "Attack roll", "Ability check"). Does not affect computation, only appears in output for readability.'
• Add error classification per pattern 'error-classification': On parse failure, return error type (invalid_syntax, unknown_modifier, out_of_range) so the LLM can decide whether to retry with a different expression or ask the user.