Zaungast provides 10 read-only tools for Teams cache inspection with consistent naming (verb_noun pattern: list_, read_, get_, search_, rank_, find_, describe_). All tools have descriptions (avg ~80 chars, within baseline 34-392). Input schemas are present and typed. However, parameter descriptions are sparse or missing entirely (e.g., 'kind' in list_conversations lacks guidance on when to filter; 'window' in rank_topics has no format spec like '7d' vs '7 days'). Output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling is minimal; no recovery guidance. All tools are read-only (low risk), but lack the polish expected of production tools.

• Document output schemas for all tools. For list_conversations, specify: {conversations: [{id: string, name: string, kind: 'chat'|'channel'|'group', lastMessageTime: ISO8601, participantCount: number}], total: number}. This enables LLMs to extract and chain IDs.
• Add format constraints to parameters. E.g., 'window' in rank_topics: 'Time window as a duration string (e.g., "7d", "30d", "1h"). Defaults to "7d".'. 'kind' in list_conversations: 'Filter by conversation type. Omit to return all types.'
• Implement error recovery guidance. E.g., if read_conversation fails with 'Conversation not found', return: 'Conversation "xyz" not found. Try list_conversations() to see available conversations, or search() to find messages by content.'
• Add pagination support. Modify list_conversations, list_events, list_calls to accept 'limit' (default 20, max 100) and 'offset' parameters. Return {items: [...], total: number, hasMore: boolean} to guide multi-page iteration.
• Clarify identifier formats in descriptions. E.g., 'conversation' parameter: 'Conversation handle or identifier (typically an email for 1:1 chats, channel name for channels, or group ID for group chats). Call list_conversations() first to discover valid identifiers.'