MCP server exposing retrieval (database read/search) and A2A (agent-to-agent) channels for the iftree document editing system. Provides tools for document structure navigation, content reading, semantic search, and agent-driven document modifications.
Three tools with detailed Chinese descriptions and complex schemas, but critical gaps in English clarity, parameter descriptions, and error handling. All tools are READ_ONLY (good for safety), but descriptions are verbose and assume domain knowledge. Schemas are present but inconsistently documented. No output schemas visible. Error recovery guidance absent.
按 library 文件夹包含层级列出已导入文档的 ASCII tree;默认文件节点显示文件名、字数和语义状态,不显示 UUID;uuid=true 时附 #docId。未导入文件不列出;(xxx字) 是该文档节点 1 的整棵子树正文合计,不是节点 1 自有正文。默认不显示摘要,includeSummary=true 时才附加摘要内容。
读取文档某地址的正文,只回正文、不带节点头。scope 选范围:subtree(默认,整棵子树正文拼接)/node(只本节点)/siblings(同父前中后三条)。必须定位到具体节点:address 传到小节(如 1-1、1-3-2)或给 nodeId——本系统的文档是检索定位用的、不是从头通读,直读根地址 1(=整篇)是反常用法,正常连读开头都该传 1-1。子树正文有 1 万字门禁(超了分层早停、只返回前几层并提示),要完整读大段得显式加大 limit 二次突破。全篇结构用 tree、定长原文窗口用 article、元信息/出处/引用用 inspect;历史版本传 at。命中过碎时读父地址或相邻地址补上下文。
查看文档结构(缩进 ASCII 树:地址 类型 标题 (子树字数))。可选 address 只看某子树,depth 限层。地址形如 1、1-3、1-3-2,是相对地址。注意字数是整棵子树合计,不是该节点自有正文。
Tool descriptions are in Chinese and assume deep domain knowledge (e.g., 'ASCII tree', 'semantic status', 'node identity piercing'). English-speaking LLMs cannot reliably parse intent. Descriptions exceed 1024 chars and bury key details in technical jargon.
Parameter descriptions are sparse or missing. 'folder' in library_index has a description, but 'includeSummary', 'includeHidden', 'uuid' lack clarity on when to use them. 'at' parameter in tree/read accepts ['string','string'] (invalid schema) with vague 'commit id/committed_at/summary tag' guidance.
No output schemas documented. LLMs cannot predict what fields to expect from library_index (does it return a tree string? structured JSON?), tree (format of ASCII output?), or read (plain text? structured object?). Breaks downstream tool chaining.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
No error handling or recovery guidance. If a docId is invalid, address malformed, or limit exceeded, what does the tool return? How should the LLM retry? 'read' mentions '1 万字门禁' (10k char gate) but provides no structured error response or guidance.
Parameter 'at' in tree and read uses invalid JSON Schema: type is ['string','string'] (array of two identical strings). Should be {type:'string'} or {enum:[...]}. This will fail schema validation in strict MCP clients.