Two tools with detailed Chinese descriptions and complex parameter schemas, but critical gaps prevent higher scoring. Browser tool has 15+ parameters with types and descriptions, but lacks output schema documentation and error recovery guidance. Terminal tool similarly lacks output schema. Neither tool follows verb_noun naming convention (should be 'control_browser', 'manage_terminal'). Descriptions are lengthy (500+ chars) but written in Chinese, making LLM optimization unclear. No evidence of idempotency hints, confirmation patterns for destructive ops, or structured error responses. Tool definitions appear directly visible in code (McpTool.kt), not inferred.

• Rename tools to verb_noun format: 'control_browser' and 'manage_terminal'. Add action parameter to tool name if needed (e.g., 'browser_navigate', 'browser_click').
• Document output schema for both tools. E.g., browser returns {tabId: string, url: string, title: string, screenshot?: string, backbone?: AccessibilityTree}. Terminal returns {tabId: string, output: string, exitCode?: number}.
• Add English descriptions (50-200 chars) alongside Chinese. E.g., 'Control a headless browser with multi-tab support. Actions: navigate, click, fill forms, take screenshots, extract text/HTML, inspect accessibility tree.'
• Implement confirmation pattern for destructive operations. Terminal 'start' with elevate=true should return {requiresConfirmation: true, reason: string} and require a separate 'confirm_elevate' call.
• Add error classification to descriptions. E.g., 'Errors: navigation_timeout (retryable), selector_not_found (user-fixable, check selector), permission_denied (fatal). On failure, return {error: string, code: string, retryable: boolean, suggestion: string}'.

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

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