Agents ✦ 90outlineCommunity tooling — calibrate on your hooks

Agent harness 督导

在回合/工具边界并行提问;harness 转向或拦截。

Who: Teams wrapping Codex/Claude Code/Pi-style agents. Steps: On turn/tool boundary ask parallel questions — skill Choice, action allow/ask/deny, stuck/slop/done Nouls → harness steers or blocks. Expected effect: cheap semantic supervisor around the agent (community cluster).

1

复制提示词(保留英文原文效果最好)。

2

在代码里实现——调用 System One / Jev,由你的程序组合答案。

3

回答它的设置问题——大多数提示词会先采访你,然后进入固定节奏。

提示词

提示词
Audience: teams wrapping Codex / Claude Code / Pi.
Steps:
1) On turn or tool boundary, snapshot recent actions + rules.
2) Parallel: skill Choice; action allow/ask/deny; Nouls stuck, slop, false_done.
3) Harness steers, warns, or blocks — worker model stays separate.
Expected effect: cheap semantic supervision (community cluster).

需要授权:typesafe-sdkagent-hooks

它是怎么搭起来的

以下说明为英文原文(来自社区作者),提示词本身建议保留英文。

Sources: pi-warden · foreman · jev-review · awesome-jev-usecases · Harness.

Cover the six-guard surface in prose: done-claim, stuck-loop, slop, irreversible, exfil, off-track — each as its own Noul/Score with code-side policy.

Sources (wave-2 deepen)

为什么它好用

Independent typed supervisor above workers; enforcement stays in the harness.

本案例收集自公开的 Jev 社区,版权归原作者所有。提示词属于参考资料:运行前请先审阅内容, 并且不要让 bot 超出你实际授予的权限行事。