Loading ToolBench
司法案例库 MCP Server — 最高法指导案例/公报案例/案例库 语义检索
Server has 8 tools with Chinese descriptions and reasonable parameter schemas, but critical gaps undermine production readiness. Tool descriptions are present but often lack clarity on WHEN to use each tool vs. alternatives (e.g., sync_now vs. fetch_release_data vs. import_workspace all modify state but lack error recovery guidance). Parameters have types and descriptions, but several tools expose implementation details (dry_run, max_pages) without explaining consequences. Output schemas are not formally documented, the code shows rendering logic (tables, cards, JSON) but no explicit schema definitions for downstream tool chaining. Error handling is minimal: no recovery guidance, no categorization of retryable vs. fatal errors. Tool composition is weak: search_cases and get_case_detail are well-separated, but the three write tools (sync_now, import_workspace, rebuild_embeddings) lack idempotency guarantees or confirmation patterns, risking data loss if an agent retries.
案例库去重与清理（标记 abolished 条目、修复 gist 串号、跨源去重），仅在 stdio 模式下可用
从 GitHub Release 下载最新数据包（cases.jsonl + embeddings.npy），用于首次初始化或数据更新
获取单个案例的完整详情（包括基本信息、规则提取、引用摘要、原文链接）
从本地 workspace 目录导入新案例（解析 YAML frontmatter + Markdown 正文），仅在 stdio 模式下可用
重建全量 embedding 索引（从 cases.jsonl 生成 embeddings.npy），仅在 stdio 模式下可用
语义检索案例库，支持关键词/自然语言查询，返回排序后的案例列表（含权威分层、引用价值、规则提取）
首次启动向导，用户选择初始化模式（release/crawl/hybrid）
触发案例库增量同步（从 rmfyalk 爬取最新案例），仅在 stdio 模式下可用（HTTP 模式为只读）
Write tools (sync_now, import_workspace, rebuild_embeddings, dedup_and_clean) lack idempotency guarantees and confirmation patterns. No dry-run feedback or rollback mechanism. Agents retrying on transient failures risk duplicate data ingestion or index corruption.
Output schemas not formally documented. Code shows rendering logic (tables, cards, JSON) but no explicit schema definitions for search_cases and get_case_detail results. Downstream tools cannot reliably extract fields like local_id, rule_text, or combined_score without parsing rendered text.
Error handling is minimal. No recovery guidance (e.g., 'Case not found. Try search_cases() first.'). No error categorization (retryable vs. user-fixable vs. fatal). Agents cannot distinguish transient failures from permanent ones.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
Three write tools (sync_now, import_workspace, rebuild_embeddings) are STDIO-only and unavailable in HTTP mode. Descriptions do not explain this limitation clearly. Agents may attempt these tools in HTTP deployments and fail silently.
setup_wizard description is vague ('首次启动向导'). Does not explain when to call it, what state it modifies, or whether it is idempotent. The choice parameter uses magic integers (1, 2, 3) instead of enums, forcing LLMs to guess valid values.