Loop Prevention

操作循环防止规范。防止 Agent 对同一操作反复执行造成资源浪费。适用于 exec 命令、工具调用、文件读写等所有可重复操作。核心规则：同一操作最多 5 次、结果不理想换策略、检测到循环立即停止。

Security Scan

The skill is an instruction-only guideline for preventing repeated operations (loops); its requirements and instructions align with its stated purpose and it requests no credentials or installs.

Purpose & Capability

Name/description match the SKILL.md: guidance to detect and stop repeated execs, tool calls, file reads/writes and API requests. No unrelated credentials, binaries, or installs are requested.

Instruction Scope

Instructions are procedural guidelines and examples (redirecting output to /tmp/out.txt, using head/tail/wc, changing params/tools). They do not ask the agent to read unrelated system files, call external endpoints, or exfiltrate data.

Install Mechanism

No install spec and no code files — instruction-only skill; nothing is written to disk by the skill itself beyond example usage of /tmp in the docs.

Credentials

No environment variables, credentials, or config paths are required. The examples reference common temporary files but do not require secrets or service tokens.

Persistence & Privilege

always is false and the skill is user-invocable; it does not request permanent presence or modify other skills or system-wide settings.

Assessment

This is a coherent, low-risk guideline for preventing repeated operations. It's just policy text — it doesn't enforce behavior by itself. Before relying on it, ensure your agent's runtime actually implements counting and stopping logic (the SKILL.md describes the policy but enforcement is up to the agent). Note the examples write to /tmp; that is normal but means temporary files may be created. Also be aware these heuristics can produce false positives (legitimate repeated retries might be stopped); test with your agent in safe environments.

防止对同一操作反复执行。适用于 exec 命令、工具调用、文件读写、API 请求等所有可重复操作。

核心规则

1. 同一操作最多 5 次 — 超限必须停止
2. 结果不理想时换策略 — 换参数、换工具、换思路，不要重复相同操作
3. 检测到循环倾向立即停止 — thinking 中出现"再试一次"、连续 2 次相同操作 → 停止并向用户说明

触发信号

信号	说明
连续 2 次相同或极相似的操作	循环开始
输出被截断后再次执行相同命令	截断≠失败，用重定向
thinking 中出现"再试一次"意图	自我循环预警
结果越来越差但还在继续	必须停止

正确做法

输出被截断

# ❌ 反复执行
git diff

# ✅ 重定向或限制
git diff > /tmp/out.txt && head -200 /tmp/out.txt
git diff --stat

工具调用失败

# ❌ 相同参数反复调用
tool_call(same_params)

# ✅ 换参数或换工具
tool_call(different_params)
other_tool(same_params)

需要完整内容

command > /tmp/out.txt
head -50 /tmp/out.txt    # 前段
tail -50 /tmp/out.txt    # 后段
wc -l /tmp/out.txt       # 总量

检查清单

问题	行动
这个操作执行过吗？	是 → 计数，超限则停止
上次结果不理想？	是 → 换策略
在想"再试一次"？	是 → 停止，换方法
连续 2 次操作相同？	是 → 循环信号，立即停止