Diagram Forge shows solid effort in parameter constraints and descriptions, but several tools lack comprehensive schema documentation and error guidance. The 6 tools vary in quality: generate_diagram and edit_diagram have complete enums and constraints, but output schemas are not documented in the code. list_templates, list_styles, and provider_health lack input schema detail. get_usage_report has good parameters but missing output specification. Error handling is present but generic, responses show basic error dict structure but lack recovery guidance or actionable next steps per the pattern:recovery-guide rubric. Tool naming is verb-forward and appropriate (generate_, edit_, list_, provider_, get_). Descriptions are adequate (80-150 chars) but could be more LLM-optimized per pattern:tool-description baseline (target 50-200 chars). The server uses FastMCP (modern), proper parameter validation (e.g., enums for diagram_type, resolution, aspect_ratio), and includes smart output_path rejection logic to prevent silent failures, this shows security awareness. However, the codebase truncates before showing tool implementations, so return schemas are inferred rather than explicitly visible.

• Add explicit return type documentation to all tools. Example for generate_diagram: 'Returns {status: str, image_path: str, image_data: bytes?, generation_id: str, cost: dict}' or similar. This lets LLMs understand what to extract from the response.
• Enhance error messages with recovery guidance. Instead of 'error: Provider not found', return 'error: Provider X not found. Call provider_health() to see available providers, or use provider="auto"'. This enables the LLM to self-correct.
• Replace imperative disclaimer text ('LEAVE AS DEFAULT') with parameter constraints. For provider: use enum ['auto', 'openai', 'gemini'] + a description 'Auto-selects the best provider for the diagram type.' For model: mark as nullable and describe 'If omitted, the server selects the optimal model for the chosen provider.'
• Document what list_templates and list_styles return. Add response structure to their descriptions: 'Returns list of template objects with fields: {name: str, description: str, diagram_types: str[], icon?: str}'. If the result could be large, add pagination: limit=50, offset=0, total_count.
• Add error classification to tool descriptions. E.g., for generate_diagram: 'If an API call fails, consider retrying with provider="auto" to switch providers. Check provider_health() if errors persist.' This guides the agent's recovery strategy.
• Clarify the behavior of nullable/optional parameters in provider_health. Document: 'If provider is omitted, returns health status for all configured providers. If specified, returns status for that provider only. Returns {status: "ok"|"error", providers: {name: string, available: bool, latency_ms?: number}}'