ai-look-up logo

ai-look-up

免费新上架

AI 抬头 | Mid-run self-review nudges for aimless AI coding agents: 检测重复命令/失败循环/零产出并注入提醒,真实本地时钟校准 AI 时间感 (ZCode / DeepSe...

访问官网
根据共同标签和关键词推荐dsh-consoledeepseek-harness-clouddshline
ai-agentagent-hooksai-lookupdeepseek-harnessdsh-pluginaiai assistantautomation

工具概览

AI 抬头 (ai-look-up) is a ZCode plugin that makes an AI agent self-review during long idle-running tasks. According to the project page, the most common waste in long tasks is "空跑" (idle running), where the same probe script is run repeatedly under different names, failed commands are retried unchanged, and no file is modified for one or two hours. The plugin uses session hooks to watch for these behaviors and, when triggered, asks the AI to pause, compare against the original goal, check whether recent steps produced anything new, and then decide to continue, change approach, or report.

The plugin registers four session hooks that observe without intercepting or blocking any tool calls: UserPromptSubmit, PostToolUse, PostToolUseFailure, and Stop. It detects idle-running features such as identical repeated commands, similar commands after normalization, long zero-output sessions, dense zero-output call sequences, and failure loops. Reminders are injected via additionalContext / agent.inject() as a structured self-review checklist.

Later versions add an independent reviewer (v0.4), goal-drift inspection and revert-edit detection (v0.5), and a third channel with an independent watcher plus concurrent state safety (v0.7). The project states that both hosts share the same detection logic, with thresholds from spec.json, and that behavior changes should be made in spec.json rather than separately in each side's code.

定价

免费

目录尚未确认完整价格档位,请到官网查看当前额度、限制及计费方式。

访问官网

价格与限制可能变化,请以官网为准,并确认币种、付款周期、最低席位数和用量上限。

应用场景

  • 团队用该插件检测 AI 智能体在长任务中是否以不同名称反复运行同一个探测脚本。
  • 团队用该插件在检测到空转行为时,让 AI 暂停并将最近的步骤与原始目标进行比较。
  • 团队用该插件通过累计连续的工具失败来识别盲目重试循环。
  • 团队用该插件在长期零产出时,于会话结束时强制进行复盘。
  • 团队用该插件注入真实的本地时间和已用时长,让 AI 依据真实时钟来估算时间。
  • 团队用该插件在空转提醒触发时,运行具有隔离上下文的独立 LLM 审查员。
  • 团队用该插件在长上下文任务中每 30 次调用定期检查目标漂移。
  • 团队用该插件通过记录每次文件修改的内容指纹来检测回退式编辑。

功能特性

AI 助手

自然语言界面

工作流自动化

用户评论

还没有评论,来分享你的使用体验吧