Six tools with complete input schemas and descriptions. Naming follows verb_noun pattern (kb_search, kb_get_doc, kb_save_*). Descriptions are detailed (150-300 chars) and explain WHEN to use each tool. However, output schemas are not formally documented, responses are plain text rather than structured JSON. Parameters lack enum constraints (e.g., doc_type accepts free-form strings instead of a closed set). Error handling is minimal; no recovery guidance for common failures (e.g., 'doc_id does not exist' returns a bare message with no next steps). Security is reasonable (no secrets in params, read-only tools marked), but no permission gates or audit trail declarations.

• Document output schemas formally. For kb_search, declare: {total: int, items: [{id: int, doc_date: string, doc_type: string, institution: string, title: string, external_id?: string, snippet?: string, ocr_required?: bool}]}. For kb_save_report/kb_save_file, return structured JSON with typed fields instead of JSON strings.
• Convert doc_type, source, and file_mode parameters to enums. E.g., doc_type: enum [announcement, research_report, minutes, article, prospectus, training, other]. This prevents hallucinated values and makes the tool self-documenting.
• Add recovery guidance to error responses. E.g., 'doc_id=999 does not exist. Call kb_search() to find documents by keyword or filter.' This guides LLMs to the next step.
• Expand kb_stats description to 100+ chars explaining what metrics are returned and when to call it (e.g., 'Returns library totals, breakdown by type/source, OCR backlog, and size. Call this to assess library health or plan ingestion.').
• Add permission declarations and audit logging. Declare scope 'write:kb' for kb_save_report/kb_save_file. Log caller ID, tool name, parameters (sanitized), timestamp, and result to stderr or a structured audit log.