Server has 6 tools with clear verb-noun naming (list_, get_, read_, send_). All tools have descriptions (avg 120 chars, within baseline 34-392). Input schemas are present with type definitions and parameter descriptions. However, output schemas are undocumented, tools return JSON strings without declaring field structure. Error handling is minimal: no recovery guidance, no categorization of retryable vs fatal errors. Missing tool annotations (readOnlyHint, destructiveHint). Composition is reasonable: tools are single-purpose, but send_and_wait combines send + poll logic, violating single-responsibility. No pagination support despite list_sessions potentially returning many results.

• Document output schemas for all tools. E.g., list_sessions should declare: returns {sessions: [{id, title, status, created_at}], total_count}. get_session_info should declare the session object structure (id, title, status, created_at, etc.).
• Add error handling with recovery hints. Wrap tool logic in try-catch; on invalid session_id, return 'Session not found. Call list_sessions() to find valid session IDs.' On timeout, return 'Request timed out. Retry with a longer timeout parameter.'
• Split send_and_wait into two tools: send_message (fire-and-forget, returns confirmation) and wait_for_session_response (polls until result event, accepts session_id and timeout). Agents can compose them as needed.
• Add tool annotations via FastMCP. Mark send_message and send_and_wait with destructiveHint=true. Mark read-only tools with readOnlyHint=true. This enables agents to reason about side effects.
• Add pagination to list_sessions: accept limit (1-100, default 20) and offset (default 0) parameters. Return {sessions: [...], total_count, has_more}. Document in description: 'Returns up to 20 sessions per call; use offset to paginate.'