The server defines 2 tools with Zod schemas and descriptions, but both exhibit significant quality gaps. The hello tool is trivial and underdescribed. The generateImage tool has a complex schema with many optional parameters, but descriptions are inconsistent in depth and quality. Chinese-language descriptions reduce clarity for international developer adoption. No input validation logic visible beyond Zod type coercion. Error handling returns generic text rather than actionable guidance. Naming is clear but lacks verb-leading patterns in some contexts. Output schema for generateImage is undocumented, LLMs do not know the structure of returned image URLs (array? string? object with metadata?). No resource or tool descriptions explain when to use each tool or why an agent should select it over alternatives.
添加即梦AI图像生成工具
添加一个简单的问候工具
hello tool lacks substantive description and is trivial (greeting only); wastes tool slot and offers no real capability to agents
generateImage output schema is not documented; code returns URLs as plain text in a 'content' array, but LLMs cannot infer whether the response contains one URL, multiple URLs, or structured image metadata (dimensions, format, generation time, etc.)
Descriptions use Chinese language (e.g. '生成图像的文本描述', '反向提示词'); this creates friction for international developers and AI model interpretability. English is the lingua franca for tool integration.
No parameter descriptions for the 'name' input to hello tool; Zod schema has .describe(null) for name, making it ambiguous what 'name' represents (user's name? nickname? identifier?)
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 43 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 31 | - | v1 |
Error handling returns generic text (e.g. '图像生成失败:未能获取图像URL') without recovery guidance; LLMs cannot determine whether to retry, check parameters, or escalate to user. No categorization of errors as retryable vs. user-fixable.
generateImage accepts model names as free-form strings (z.string().optional()) rather than enums; description documents valid values (jimeng-4.0, jimeng-3.0, etc.) but LLM can hallucinate invalid models like 'jimeng-5.0' or 'jimeng-pro'. Enums prevent this.
sample_strength parameter lacks validation bounds; description says 'range 0-1' but Zod schema is z.number().optional().default(0.5) with no .min().max() constraints. LLM can pass 2.5 or -1.
width and height parameters have no constraints; LLM can request 10000×10000 pixel images or 1×1, causing API failures or excessive latency. Numeric ranges (e.g. 256 - 2048) should be enforced.
No tool annotation hints (readOnlyHint, destructiveHint, idempotentHint); generateImage has WRITE risk but tool registration lacks explicit marking. Clients cannot determine safety classification without reading descriptions.
Resources exposed (greeting://{name}, info://server, jimeng-ai://info) lack practical value and documentation; they are static text blobs that could be part of README instead. Wastes resource slots.