Critical gaps across all dimensions. Tool descriptions are in Russian and lack LLM-optimized English guidance. Only 2 of 4 tools have input schemas; 2 tools (layout_snapshot, selected_element) have schemas but descriptions are vague about when/why to call them. No output schemas documented. Error handling absent. Parameter descriptions exist but are minimal (10-30 chars). No enum constraints despite tools accepting specific values (role param in reply_in_window). Naming is non-standard (underscores in Russian context, no clear verb-noun pattern). No security considerations for write operations.

• Translate all tool descriptions to English. Provide 50-200 character descriptions explaining WHAT each tool does, WHEN to call it, and WHAT it returns. Example: 'Get the current DOM tree snapshot with node IDs and bounding boxes. Call this first to understand the page layout before selecting elements. Returns tree with max depth parameter.'
• Add explicit output schemas for all tools. Document the structure of layout_snapshot (e.g., {nodes: [{id, tag, text, rect, anchors}], depth}), selected_element (e.g., {id, box, parents, editable}), pending_requests (e.g., {requests: [{id, comment, artifacts, measurements}]}), and reply_in_window (e.g., {success, messageId}).
• Add descriptions to all parameters. For reply_in_window.role, explain: 'assistant = response to user; system = internal note visible only in window'. For pending_requests.includeConsumed, explain: 'If true, return already-processed requests; if false, only new ones'.
• Add enum descriptions and validation. For maxDepth in layout_snapshot, add description: 'Tree depth limit (1-30). Higher values return more detail but larger payloads. Default 8 balances detail and performance.' For role, explain each enum value.
• Document error scenarios and recovery paths. Example: 'If no element is selected, reply_in_window returns error: "No element selected. Call selected_element first to verify a selection exists." Retryable: yes.'