Loading ToolBench
缙云文采（Sovena）：Zotero → Markdown 语义文献包 → 向量检索，各学科通用的本地文献流处理与 MCP 服务（含任意文件夹 OCR/索引的 adhoc 临时任务）
Sovena has 16 well-named tools with consistent verb_noun patterns (zotero_*, sovena_*). Descriptions are present and moderately detailed (avg ~150 chars), explaining WHAT each tool does and WHEN to use it. However, parameter descriptions are sparse or missing for several tools (e.g., sovena_adhoc_process lacks descriptions for 'path' and 'name'). Output schemas are not documented, responses are free-text strings rather than structured JSON. Error handling is minimal; most tools return plain text without recovery guidance. Security is reasonable (no exposed secrets in params), but no tool annotations (readOnlyHint/destructiveHint) are present despite clear risk distinctions (WRITE vs READ_ONLY tools).
已有资料包列表。
任意文件/文件夹 → OCR+向量索引。
取消一个排队/运行中的任务。
连通性诊断（Zotero/embedding/目录权限/OCR）。
找与某条 Zotero 文献语义相似的其他文献块（基于向量索引）。
查询任务进度与最近日志。
列出全部任务（最近 100 个）。
查看某分类的语义文献包清单（条目、转换路由、页数、状态）。
Output schemas not documented. All tools return free-text strings (e.g., zotero_collections returns '\n'.join(lines)) rather than structured JSON. LLMs cannot parse results reliably or chain downstream tool calls without explicit schema documentation.
Parameter descriptions missing or incomplete. sovena_adhoc_process has 'path' and 'name' params with no descriptions. zotero_search 'q' param lacks guidance on what 'full-text keyword' means. LLMs cannot infer valid input formats.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
将某 Zotero 子分类一键做好 AI 调用准备（采集→转换→语义包→向量索引）。默认增量：只处理新增或修改过的条目（依据 Zotero version 与附件指纹）。
读取文献 markdown 全文。
自然语言语义检索已准备的文献（含 adhoc 资料包，collection 形如 adhoc:书库名）。返回带原书页码（page 字段）的命中片段，支持跨语言（中文查询命中外文文献）。
系统资源状态。
读取某条文献 PDF 上的标注（高亮/批注）。可按颜色（如 #ffd400）与类型过滤。
列出 Zotero 文库全部分类（层级缩进）及各分类语义包准备状态。
获取单条 Zotero 条目的完整信息：元数据、附件、笔记。
按元数据搜索 Zotero 条目（q 为全文关键词，其余为精确过滤）。返回条目 key、标题、作者、年份、附件情况。拿到 item_key 后可用 zotero_item 看详情，或用 sovena_read_item 读已准备的全文。
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Tools are marked with Risk labels (READ_ONLY, WRITE, REVERSIBLE) in metadata but not exposed to the MCP client. Agents cannot distinguish safe reads from destructive writes without explicit annotations.
Error handling lacks recovery guidance. Tools return plain text errors (e.g., 'task not found') without suggesting next steps. No categorization of retryable vs. fatal errors. LLMs cannot self-correct or plan recovery.
Pagination not implemented. zotero_search and sovena_search accept 'limit' but no offset/cursor. Large result sets are truncated silently. No indication of total count or whether more results exist.