Three tools with basic schemas and minimal descriptions. All tools have input schemas with required parameters properly defined (hospital_id, sku, units). However, descriptions are extremely terse (10-30 chars), falling well below the 194-char baseline for production tools. No parameter-level descriptions exist, LLMs cannot infer what 'hospital_id' or 'sku' represent without additional context. Tool annotations (destructiveHint/readOnlyHint) are present in the wire format, which is good. Error handling in the server catches exceptions and returns isError=true, but tool-level error guidance is absent. No output schema documentation. The server code is well-structured and implements proper version negotiation and JSON-RPC error handling, but the tool definitions themselves lack the depth needed for reliable LLM selection and execution.

• Expand tool descriptions to 100-200 characters. Example: 'get_inventory: Returns the current days of inventory cover for a hospital. Use this to check stock levels before deciding whether to raise a purchase order. Requires a valid hospital_id.'
• Add parameter descriptions in the input schema. For each parameter, add a 'description' field explaining its meaning, format, and constraints. Example: 'hospital_id: The unique identifier for the hospital (string, 1-50 chars, alphanumeric).'
• Document output schemas. Add a 'returns' section to each tool description or a separate schema document. Example: 'get_inventory returns {days_of_cover: number, last_updated: ISO8601 string, status: "ok"|"low"|"critical"}'.
• Add validation and error guidance. In tool descriptions, state expected ranges and error recovery steps. Example: 'units must be 1-10000. If the PO fails, call get_inventory to verify current stock and retry.'
• Implement a dry-run or confirmation pattern for raise_purchase_order. Add an optional 'dry_run: boolean' parameter or a separate 'confirm_purchase_order' tool to prevent accidental destructive calls.

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

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained