Loading ToolBench
Antigravity global MCP plugin for persistent memory and context relay across agent sessions
Server has 7 tools with basic schemas and descriptions, but significant quality gaps. All tools have descriptions (positive), but most are under 100 chars and lack actionable context. Schemas are present but minimal, parameters have types and descriptions, but no enums, ranges, or validation guidance. No error handling patterns visible (no recovery guidance, no categorization). Tool naming is reasonable (verb_noun style) but some names are vague ('auto_summarize_and_save' combines multiple concerns). No tool annotations (readOnlyHint, destructiveHint). Output schemas are not documented. Security concern: API keys loaded from config file but no validation that file permissions are correct (chmod 600 not enforced). Composition is weak, tools operate on a single SQLite database with no pagination, no batch operations, and no result limits documented.
【省 Token 专用】接收原始对话日志，由 Kimi 小模型自动压缩摘要后保存快照。主模型无需自行生成摘要，直接传原始日志即可。
获取项目当前状态。
新开对话时，读取指定角色的快照，恢复上下文。
检索历史记忆。可按关键词过滤。
记录关键决策。任何会话都能检索到。
对话过长需要关闭重开时，保存当前角色的快照。
更新项目状态。
Tool 'auto_summarize_and_save' combines two responsibilities (summarize + save). Should split into separate tools: 'summarize_chat_logs' and 'save_context'. Violates single-responsibility principle.
No output schemas documented for any tool. LLMs cannot predict return structure or plan downstream calls. All tools return free-text strings instead of structured objects.
No error handling guidance. Tools return bare success/failure strings without recovery hints. E.g., 'load_context' returns '⚠️ 未找到...' but does not suggest next steps (e.g., 'Call save_context first').
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | <=2025-11-25 | v2 |
No pagination or result limits. 'recall_memories' accepts a 'limit' parameter but no documentation of default, max, or total count in response. Large result sets will blow context window.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot distinguish safe reads from destructive writes. 'update_project_status' and 'record_decision' should be marked destructiveHint=true.