Server has 2 tools with explicit schemas and descriptions visible in source. deploy_vm has a complete input schema with 7 parameters, all typed and described. restart_service has a complete input schema with 4 parameters, all typed and described. However, there are significant gaps: (1) No output schemas documented for either tool, the functions return plain strings, not structured objects. This violates pattern:tool and makes it impossible for LLMs to plan downstream calls or extract structured data. (2) No error handling strategy documented, no guidance on retryability, user-fixable vs fatal errors, or recovery paths. (3) No tool annotations (destructiveHint, idempotentHint, readOnlyHint) despite deploy_vm being explicitly irreversible and restart_service being reversible. (4) Descriptions are adequate but generic, deploy_vm's description (65 chars) explains WHAT but not WHEN to use it relative to other VMs or alternatives. (5) Admin password parameter accepts free-form strings; the constraint '(min 12 chars, must include uppercase, lowercase, number, special char)' is only in the description text, not enforced via regex pattern or enum. LLMs cannot reliably read prose constraints. (6) No pagination or result limits documented, though these are single-resource operations so less critical. (7) Credentials are hardcoded in the class initialization, environment variables are read (load_dotenv), but the server does not validate or document scope/permissions required.

• Define output schemas as typed objects. For deploy_vm, return {status: 'success'|'error', vm_id: string, public_ip: string, fqdn: string, admin_username: string, estimated_cost_per_month: number} or similar. This enables chaining and structured response handling.
• Add error classification: return {error: string, category: 'permission_denied'|'invalid_input'|'resource_not_found'|'transient'|'unknown', recovery: string} for all error cases. Example: {error: 'Resource group already exists', category: 'invalid_input', recovery: 'Use a unique resource group name or omit to auto-create.'}
• Add tool annotations: @app.tool(destructiveHint=True) for deploy_vm; @app.tool(idempotentHint=True) for restart_service. These guide LLM behavior on confirmation and retry policies.
• Enforce password validation at tool entry: validate regex before any Azure API calls. Return clear error: 'Invalid admin_password: must be 12+ chars, with uppercase, lowercase, digit, and special char (@$!%*?&).'
• Replace prose examples with enums in schema. For location, define enum: ['eastus', 'westus2', 'centralindia', 'northcentralus', ...] (full list). For os_type, already enforced in code but make enum explicit in schema: ['linux', 'windows'].
• Add dry_run parameter to deploy_vm (default False) that returns the configuration and estimated costs without deploying. This lets users preview and validate.
• Document idempotency guarantees. If deploy_vm is idempotent server-side, state: 'Idempotent: calling twice with the same resource_group/vm_name returns the existing VM, no duplicate created.' If not, implement client-side deduplication via idempotency keys.

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

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