Jev 101
开始指南
搜索
中文 EN

使用案例 tool-call-guardrails

Agents ✦ 94outlineAuthor-reported metrics — re-measure on your hooks

工具调用 / 越狱护栏

每个危害一类 Noul；按危害设阈值 allow/ask/block。

Who: Agent harness authors. Steps: (1) State = tool name, args summary, user utterance, rules. (2) Prefer one Noul per hazard (destructiveness, exfil, scope creep) — not one vague “is this dangerous?”. (3) Code maps probabilities → allow / ask / block with per-hazard thresholds. Expected effect: fewer false negatives vs a single broad question (firewall/pi-warden authors report this pattern — cite as author-reported).

来源：AnshChoudhary/typesafe-ai-firewall ↗ 文档 / 来源 ↗
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: agent harness authors.
Steps:
1) Capture proposed tool + args summary + utterance + rules.
2) Parallel Nouls — one per hazard (destructiveness, exfil, scope_creep, injection_in_tool_result).
3) Code: per-hazard thresholds → allow / ask user / block.
Expected effect: stronger screening than one vague danger question (author-reported in community firewall studies).
Do not execute tools from inside the model.

需要授权：typesafe-sdkagent-runtime

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。

Insert a pre-execution gate. Sources: llm_guardrails cookbook · typesafe-ai-firewall · pi-warden · pi-jev. Measured claims (paired runs / hard-negative rates) are author-reported.

Prefer one Noul per hazard (exfil, irreversible, jailbreak, …) with per-threshold tables in code rather than one vague “is this safe?” question. Multi-hazard firewalls are author-reported improvements over single checks.

Sources (wave-2 deepen)

为什么它好用

Guardrails are cheap relative to the tool call they protect; probabilities beat essays.

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。