Single tool 'plan_browser_actions' has critical definition gaps. Input schema is visible but incomplete: 'goal' and 'elements' parameters lack proper type definitions in the MCP registration (tools/list response shows no schema at all). Description is present (67 chars) but generic. No output schema documented. Tool definition is inferred from client.py rather than explicitly registered in MCP handler.

• Add explicit inputSchema to tools/list response: {"type": "object", "properties": {"goal": {"type": "string", "description": "Natural language goal (e.g., 'Add product to cart and checkout'). Required."}, "elements": {"type": "array", "items": {"type": "object"}, "description": "Optional DOM element list with tag, id, class, text attributes. If omitted, defaults to example e-commerce elements."}}, "required": ["goal"]}
• Expand tool description to 100-150 chars: 'Generate resilient browser action sequences from natural language goals. Returns step-by-step action plan with CSS and XPath selectors. Use when you need to automate multi-step web interactions (search, click, navigate).'
• Document output schema in description or as a separate field: action_sequence is array of {step, action_type, selector, fallback_xpath, description}; resilience_grade indicates dual-selector fallback support.
• Add input validation and error messages: if goal is empty, return 'Error: goal parameter required. Provide a natural language description of browser actions (e.g., "Search for product and add to cart").' If elements array is malformed, return 'Error: elements must be array of objects with tag, id, class, text properties.'
• Consider splitting tool into two: plan_browser_actions (returns action sequence) and validate_browser_plan (dry-run check). Agents can then preview before execution.