Server has 7 tools with complete input schemas and descriptions in Vietnamese. Naming follows verb_noun pattern (search_, get_, list_, add_). However, descriptions are domain-specific and lengthy (200-300 chars), making them less LLM-optimized. Parameter descriptions are present but terse. Output schemas are not explicitly documented, responses are inferred from code. Error handling is minimal; add_decision returns a success dict but no guidance on failure modes. No tool annotations (readOnlyHint/destructiveHint) despite clear read/write distinction. Missing pagination support in search_decisions despite accepting a limit parameter.

• Add explicit output schema documentation to each tool. Example: search_decisions returns [{"id": string, "title": string, "scope": string, "status": string, "date": string}]. Use JSON Schema in tool definition.
• Add tool annotations: mark add_decision with destructiveHint=true; mark all read tools with readOnlyHint=true. This enables LLM safety reasoning.
• Translate descriptions to English or provide bilingual descriptions. Vietnamese-only descriptions reduce LLM comprehension, especially for non-Vietnamese-trained models.
• Add offset/limit or cursor-based pagination to search_decisions. Return {"results": [...], "total": N, "offset": M, "limit": L} to enable iteration.
• Standardize error responses. All tools should return {"error": "<message>", "recovery": "<next step>"} on failure, not silent nulls or inconsistent formats.