Server has 4 tools with complete input schemas and descriptions, but quality is uneven. Tool names follow verb_noun convention (list_models, health_check, simple_chat, chat_completion). Descriptions are present but brief (34-89 chars, below the 50-200 char LLM-optimized baseline). Parameters have types and descriptions, but lack constraints (enums, ranges, patterns). No output schema documentation. Error handling returns structured responses but lacks recovery guidance. Missing tool annotations (readOnlyHint, destructiveHint). Overall: functional but below production baseline.

• Add JSON Schema output definitions for all tools. Document the structure of {'ok': bool, 'model': str, 'text': str, 'error'?: str, 'status'?: int, 'raw'?: object}.
• Add parameter constraints: model as enum(['DeepSeek-R1-671B', 'DeepSeek-R1-Distill-32B']), temperature as number with minimum 0, maximum 2, repetition_penalty as number with minimum 1, maximum 2, timeout as number with minimum 1, maximum 600.
• Expand descriptions to 80-150 chars. For simple_chat, add: 'Single-turn chat without history. Use chat_completion for multi-turn conversations. Supports streaming responses.'
• Add tool annotations via FastMCP: @mcp.tool(annotations={'readOnlyHint': True}) for all tools since they do not modify state.
• Enhance error messages with recovery guidance. E.g., 'API key expired (token valid ~5 hours). Refresh TSINGHUA_DEEPSEEK_API_KEY and retry.' or 'Model not found. Available: DeepSeek-R1-671B, DeepSeek-R1-Distill-32B'.

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

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