Loading ToolBench
A Zettelkasten knowledge management system providing atomic note-taking, bi-directional linking, full-text search, knowledge graph analysis, and human-in-the-loop review/feedback workflows for OpenClaw agents.
The server provides 34 tools with reasonable schema structures and descriptions in Chinese. However, there are significant gaps in LLM-optimization, parameter descriptions, output schema documentation, and error handling. Most tool descriptions are under 100 characters (below the 194-char baseline for A-tier tools). Many parameters lack type information in visible schema definitions, and no output schemas are documented. The tools are read-write split appropriately, but descriptions do not explain when to use a tool vs. alternatives, prerequisites, or recovery paths. Naming is consistent (zk_ prefix) but generic verbs like 'get' and 'run' dominate. Tool composition has concerning gaps: tools return unspecified structures, idempotence is unclear, and chaining IDs are not confirmed in responses.
分析反馈趋势
归档笔记
创建笔记（含置信度路由）
发现知识网络中的意外关联（Serendipity Engine）
批量蒸馏记忆（夜间批处理）
导出高质量样本
查找路径（两张卡片之间的最短路径）
获取僵尸笔记
No output schemas documented. Tools return unspecified structures, forcing LLMs to infer response shapes from context or prior examples. This violates the pattern:tool-description and pattern:response-shaper principles and increases hallucination risk.
Tool descriptions are too short and lack LLM-optimization guidance. Average length ~50 chars (baseline for A-tier is 194 chars). Descriptions do not explain WHEN to use this tool vs. similar ones, WHAT data it requires, or WHAT the LLM should do with the result. Examples: 'zk_glow_ranking' (51 chars), 'zk_find_zombies' (45 chars), 'zk_knowledge_heatmap' (40 chars).
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | <=2025-11-25 | v2 |
获取指定类型的活动提示词
获取归档历史
获取历史审计报告列表
获取反向链接（指向该笔记的链接）
获取样本策划统计
获取反馈统计
获取 Inbox 待审核队列
获取 MOC 生成统计
获取单条笔记
获取提示词效果统计
获取审核面板状态（待审核项目列表和统计）
获取审核统计信息
获取意外发现统计
获取发光度排行
生成知识库健康度审计报告
知识热力图
知识图谱
审核笔记（批准/拒绝/改进）
运行 CEQRC 工作流（深度内化）
扫描知识社区并生成结构笔记（MOC）建议
搜索已归档笔记
搜索 Zettelkasten 笔记
提交反馈
提交审核决定
恢复归档笔记
更新笔记
Many parameter descriptions are missing or trivial. The rubric requires every parameter to have a non-empty description explaining what it controls and what values are valid. Examples: 'zk_get_review_panel' has no input properties documented; 'zk_get_feedback_stats' lacks parameter detail; 'zk_get_moc_stats' and similar stats tools provide empty input schemas without explanation.
No error handling or recovery guidance in tool definitions. Descriptions do not explain what errors are possible, whether they are retryable, or what the LLM should do next. Violates pattern:recovery-guide and pattern:error-classification. LLMs have no guidance on handling failures.
Destructive operations (archive_note, unarchive_note, delete via review, create/update note) lack confirmation or dry-run patterns. Violates pattern:confirmation-request. No safeguards prevent an LLM from accidentally archiving or modifying many notes in a loop.
Tool names are generic and do not reliably convey intent. Many use 'get_', 'run_', and 'submit_', acceptable but undifferentiated. No names start with action verbs that hint at the operation's specificity. For example, 'zk_glow_ranking' does not indicate whether it returns notes sorted by 'glow' (a custom metric) or something else. LLMs may conflate similar tools.
No documented output schema means tools cannot confirm they return the IDs and references that downstream tools need. For example, zk_search_notes should return note_id so that zk_get_note, zk_archive_note, or zk_review_note can use it in a chain. Without documented outputs, LLMs cannot plan multi-step sequences reliably. Violates pattern:tool-chain and mxe:include-chaining-ids.
Pagination is unclear or absent in discovery tools. zk_search_notes accepts 'limit' but no 'offset' or 'cursor' parameter. zk_get_inbox_queue, zk_glow_ranking, and others accept 'limit' only. Without offset/cursor and a total count or next_cursor, LLMs cannot reliably paginate large result sets. Violates pattern:paginated-result.
Tool composition is unclear. Multiple tools query review state (zk_get_review_panel, zk_get_review_stats), feedback (zk_get_feedback_stats, zk_analyze_feedback_trends), and prompts (zk_get_prompt_stats). LLMs cannot easily determine which to call or how they differ. No indication of whether they are idempotent or whether calling them multiple times has side effects. Violates pattern:idempotent-operation and pattern:tool composition guidance.
No tool accepts human-friendly identifiers. All tools require IDs (noteId, targetId) rather than titles or names. This forces agents to first call search_* to find an ID, then use it in the actual operation. Violates mxe:natural-identifiers and increases latency and token use.