The server provides 16 tools with varying quality. Strengths: tool names follow verb_noun patterns and are mostly clear; descriptions are present for all tools and explain purpose. Weaknesses: descriptions are written in Chinese, which limits LLM comprehension outside Chinese-language models; input schemas are visible but lack type constraints and enums; output schemas are NOT documented (tools return untyped strings or raw data structures); error handling is minimal, most errors are simple string messages without recovery guidance or categorization; no pagination support despite query tools; no batch operations; critical missing pieces include documented return types, parameter constraints (enums for difficulty/risk levels), and actionable error messages.

• Add output type annotations to all tools. Define what fields each returns (e.g., 'Returns: List[{id: int, name: str, version: str, is_deleted: bool}]'). Use JSON Schema to document response structure.
• Provide English translations of all tool and parameter descriptions. Maintain the Chinese versions as secondary documentation if desired, but ensure the primary MCP interface is accessible to English-language LLMs.
• Add enum constraints to 'create_question' difficulty parameter. Define as enum: ['easy', 'medium', 'hard']. Validate input server-side and return 'Invalid difficulty: got "<value>", must be one of: easy, medium, hard'.
• Implement pagination for 'query_all_programming_languages' and 'query_all_tags'. Add parameters: limit (default 20, max 100), offset (default 0). Return {data: [...], total: N, next_offset: M}. Document the limit in the tool description.
• Enhance error handling with structured responses. Instead of 'permission denied', return {error: 'PERMISSION_DENIED', message: 'You do not have write access to question 123', recoverable: false}. Guide LLMs on when to retry.
• Fix tool composition. Ensure 'query_question_info' returns tag_id alongside tag name, so agents can use the IDs with other tools. Alternatively, provide a 'list_tags' tool that returns (id, name, score) tuples.
• Add parameter validation hints to descriptions. E.g., 'question_id (integer, ≥1): The unique ID of the question in the database. Must be a valid question owned by the current user.'
• Implement idempotency guarantees or add dry-run support for create/update tools. E.g., add 'dry_run: bool = false' parameter to 'create_question' so agents can preview the result before committing.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry