Single tool server with a clear domain focus (RappelConso product recalls). Tool name is reasonably action-oriented ('rappel-conso-tool'), description is present and adequate (87 chars), and input schema is properly defined with Zod. However, the tool name deviates from standard verb_noun conventions (missing a leading action verb like 'get_' or 'search_'), and the output schema is not explicitly documented in the tool definition. Parameter descriptions are present but minimal. No error guidance for LLMs, and destructive operations are not marked with tool annotations.

• Rename tool from 'rappel-conso-tool' to 'get_product_recalls' or 'search_recalls' to follow verb_noun convention and improve LLM tool selection clarity.
• Add explicit output schema documentation to the @Tool decorator. Define the RappelConsoResponse structure (total_count: number, results: array of recall objects with key fields) so LLMs know what to expect.
• Enhance error handling to provide actionable recovery guidance. Example: if query parse fails, return 'Invalid query syntax. Provide a SQL WHERE clause like: categorie_produit like \'jambon\'. Available fields: [list]'. Categorize errors as retryable (API timeout) vs user-fixable (bad query).
• Add readOnlyHint: true to @Tool decorator to signal that this tool only reads data and is safe to call repeatedly.
• Document the 'query' parameter format more explicitly: specify that it expects SQL WHERE clause syntax, list 3-5 example valid queries, and note which fields are searchable.
• Add a 'fields' parameter (optional array of strings) to allow LLMs to request only specific fields in results, reducing token overhead for large datasets.
• Implement pagination explicitly: add 'offset' parameter (default 0) alongside 'limit', and return 'offset' and 'has_more' in the response so LLMs can iterate through large datasets.

Inferred effective spec: 2026-07-28+.

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