Two tools with partial schemas and descriptions. caspian_setup_credentials has a well-structured enum-based input schema with clear parameter descriptions (mode, apiKey, baseUrl all documented). caspian_reply_email has a simpler schema (body required, messageId optional) with adequate descriptions. However, both descriptions are verbose (200+ chars) and lack clear WHEN/WHY guidance for LLM selection. No output schemas documented for either tool. No error handling guidance provided. Naming is verb-noun compliant but descriptions read more like user docs than LLM-optimized prompts. Missing idempotency hints, retry guidance, and recovery paths.

• Document output schemas for both tools. caspian_setup_credentials should return {status: 'configured'|'sandbox_created'|'error', message: string, configPath?: string}. caspian_reply_email should return {messageId: string, threadId: string, channel: 'email'|'telegram', timestamp: ISO8601}.
• Rewrite descriptions as LLM-optimized prompts: 'Set up Caspian API credentials. Use mode=status to check, mode=sandbox for instant free key, or mode=paste with your dashboard API key. Writes to ~/.config/opencode/caspian.env.' (80 chars vs 200+).
• Move apiKey from tool parameter to server-side secret injection. Accept only mode parameter; retrieve CASPIAN_API_KEY from environment or vault inside the tool implementation.
• Add error handling guidance: 'If setup fails with permission error, check ~/.config/opencode/ is writable. If API key invalid, verify at https://dashboard.trycaspianai.com.'
• Add idempotency and confirmation hints to caspian_reply_email: 'This tool sends an email reply, action is irreversible. Confirm before sending to prevent duplicates.'