8 tools with explicit schemas and descriptions, but significant gaps in parameter documentation and output schema clarity. Tool names follow verb_noun convention (comet_connect, comet_ask, comet_poll, comet_stop, comet_screenshot, comet_tabs, comet_mode, comet_upload). Descriptions range 34 - 392 chars (within baseline 10 - 1024). However, 5 of 8 tools have parameters with missing or incomplete descriptions (comet_tabs: 'domain' and 'tabId' lack context on when to use each; comet_mode: 'mode' enum is documented but no guidance on mode selection criteria). No output schemas documented for any tool, LLMs cannot predict response structure. Error handling absent; no recovery guidance. Security concern: comet_upload accepts arbitrary file paths without validation or sandboxing hints.

• Document output schema for each tool. Example for comet_ask: {type: 'object', properties: {response: {type: 'string'}, taskId: {type: 'string'}, completedAt: {type: 'string', format: 'date-time'}}}. This enables LLMs to chain tools and extract required fields.
• Clarify comet_tabs parameter relationships: 'action=list ignores domain/tabId. action=switch/close: if both domain and tabId provided, tabId takes precedence. domain matches first tab with matching domain.'
• Add validation and security guidance to comet_upload: 'filePath must be absolute and within user home directory. Supports: .jpg, .png, .pdf, .docx (max 50MB). Returns error if file not found or type unsupported.'
• Add error handling to all tools. Example: 'If comet_ask times out after 120s, call comet_stop() to cancel, then retry with shorter timeout or simpler prompt. If comet_connect fails, ensure Perplexity is installed and port 9223 is available.'
• Expand comet_poll description: 'Returns {taskId, status: pending|running|completed|failed, progress: 0-100, currentStep: string, elapsedMs: number}. Poll every 2-5s. If status=failed, call comet_stop() and retry with comet_ask.'