23 tools with consistent naming (verb_noun pattern) and basic descriptions. All tools have input schemas with typed parameters and descriptions. However, descriptions are terse (avg ~40 chars, well below the 194-char baseline), lack WHEN/WHY context, and omit error recovery guidance. Output schemas are undocumented, LLMs cannot plan downstream calls. No tool annotations (readOnlyHint/destructiveHint). Streaming tools (stream_draft_findings, stream_draft_synthesis, stream_chat_turn, stream_external_query_draft) lack pagination/limit guidance. Parameter descriptions are minimal; many lack format/constraint details. No idempotency or confirmation patterns for destructive ops (create_project, import_report, register_finding, update_synthesis).

• Expand tool descriptions to 100-200 chars. Include WHAT (action), WHEN (use case), and WHAT NEXT (expected output). E.g., 'Create a new research project to organize findings and reports. Returns project_id for use in subsequent calls.'
• Document output schemas for all tools. Specify returned fields, types, and which IDs can be chained to other tools. E.g., create_project should document that it returns {project_id, name, created_at} so LLMs know project_id is valid for get_project_overview.
• Add tool annotations: mark list_* and get_* tools with readOnlyHint=true; mark create_*, update_*, delete_*, import_*, register_* with destructiveHint=true; mark idempotent ops with idempotentHint=true.
• Add explicit bounds to numeric parameters: maxItems (1-1000, default 20), confidence (0.0-1.0), limit (1-100). Document defaults and caps in parameter descriptions.
• Add error recovery guidance to descriptions. E.g., 'If project not found, call list_projects() to discover available projects.' This guides LLM retry logic.