Critical failures across all dimensions. Tool definitions exist in src/inventory.js but lack visible input schemas, parameter descriptions, and output documentation. Descriptions are minimal (10-30 chars), far below the 50-200 char baseline for LLM-optimized tools. No parameter type definitions visible. No error handling guidance. No idempotency or confirmation patterns for destructive operations (delete_*, restore_config_backup, restart_ha). The server exposes 72 tools but provides almost no structured metadata to guide LLM selection or parameter binding.

• Add explicit JSON Schema definitions for all 72 tools. Each tool must declare inputSchema with type, properties, required array, and per-parameter descriptions. Example: {type: 'object', properties: {entity_id: {type: 'string', description: 'Home Assistant entity ID (e.g., light.living_room)'}}, required: ['entity_id']}.
• Expand all descriptions to 50-200 characters. Current descriptions (10-30 chars) are too terse. Include WHAT the tool does, WHEN to use it, and any prerequisites. Example: 'Get the current state and attributes of a Home Assistant entity. Use this to check if a light is on, a sensor's value, or any entity's status. Requires a valid entity_id.'
• Add confirmation/dry-run patterns for all 14 destructive tools (delete_*, restore_config_backup, restart_ha, batch_edit_config_files, fire_event). Implement a --confirm flag or a separate preview_* tool that shows what would be deleted before executing. Require explicit user approval via the CLI --yes flag (already present in bin/hass.js) or MCP confirmation flow.
• Document output schemas for all tools. Add a 'returns' section to each tool description or a separate schema field. Example for get_state: 'Returns {state: string, attributes: object, last_changed: ISO8601, last_updated: ISO8601}'. For list_* tools, document pagination (limit, offset, total_count).
• Add error handling guidance. When a tool fails, return actionable messages. Example: 'Entity "light.xyz" not found. Available entities: light.living_room, light.bedroom. Call list_entities() to discover all entities.' This enables LLMs to self-correct without extra discovery calls.