Smelt has 15 tools with READ_ONLY risk profile, but critical gaps in schema visibility and parameter documentation. The primary `convert` tool has a broad, helpful description (>200 chars), but most specialized tools (pdf_to_text, docx_to_text, etc.) lack visible input schemas in the provided source. Tool names follow verb_noun convention (good), but parameter descriptions and type definitions are not visible in the code excerpts. No evidence of enum constraints, validation guidance, or output schema documentation. Error handling and recovery guidance are absent. The server exposes optional specialized tools via a config flag, which is reasonable composition, but the lack of visible schema definitions for these tools prevents confident scoring above 50.

• Add explicit JSON Schema definitions for all tool parameters. Include type, description, enum constraints (e.g., output_format: ['markdown', 'xml', 'json', 'doctags']), and min/max for numeric params. Ensure every parameter has a 50-150 char description explaining its purpose and valid values.
• Expand descriptions for specialized tools (pdf_to_text, etc.) to 100-200 chars. Explain when to use each vs. the primary `convert` tool. E.g., 'Use pdf_to_text when you need explicit control over PDF backend selection (mineru, docling, pymupdf). For automatic backend selection, use convert instead.'
• Document output schemas for all tools. Specify the structure of returned text, metadata, and any pagination fields. E.g., convert returns {text: string, metadata: {page_count: int, backend: string, ...}, conversion_id: string}.
• Add error handling guidance to tool descriptions. E.g., 'If the file is not found, the tool returns an error with the file path. If a backend is unavailable, it falls back to the next priority backend. If all backends fail, the error message lists which backends were attempted.'
• Implement and document pagination for list_backends and list_formats. Add limit and offset parameters, return a total_count field, and cap results at 50 items per page. Document this in tool descriptions.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry