🔄 ClawBrain Smart Retry — 技能工具
v1.2.0v1.2 错误恢复：三级恢复策略 + 双模型共识 + 循环检测 + 降级告知 + 长对话恢复。出错自动换方案，不重复同样的错误。
1· 139·0 当前·0 累计
by 安全扫描
OpenClaw
可疑
medium confidenceThe skill's description matches an error-recovery helper, but its runtime instructions implicitly permit filesystem access, installing packages, and contacting model/memory backends without declaring required credentials or boundaries — this mismatch is concerning.
评估建议
This skill appears to be an instruction-only helper for automatic retries and fallbacks, which can legitimately need to run commands, query health endpoints, and access memory stores — but the SKILL.md omits critical details. Before installing, ask the publisher: (1) which endpoints and credentials (if any) the skill will contact/use, (2) what exact commands it might run (which package managers, whether sudo could be invoked), and (3) whether filesystem reads/writes or package installs will be s...详细分析 ▾
ℹ 用途与能力
The claimed purpose (automatic retry, fallback models, loop detection, recovery from truncation) plausibly requires checking files, running replacement commands, querying backend health, and reading memory/knowledge-graph state. However, the SKILL.md assumes access to model orchestration, memory services, and HTTP health endpoints without declaring any required credentials, endpoints, or config paths — an omission that reduces transparency.
⚠ 指令范围
Instructions explicitly suggest running shell actions (ls, install dependencies, try alternate commands), injecting recovery directives after repeated failures, querying GET /v1/health/backends, and restoring context from a knowledge graph/memory service. Those are broad actions that involve filesystem, package managers, and network calls; the skill gives no constraints or limits and could result in unintended file reads/writes, package installs, or outbound traffic.
ℹ 安装机制
There is no install spec (lowest install risk). Still, the runtime guidance tells the agent to install missing dependencies at runtime — a behavioral mismatch: the skill expects to perform installs but doesn't declare how or request permission/constraints for doing so.
⚠ 凭证需求
The skill does not request any environment variables, credentials, or config paths, yet it presumes access to model fallback chains, backend health endpoints, and memory/knowledge-graph services that typically require credentials or scoped config. Absence of declared secrets or endpoints is an incoherence and a potential security/privacy risk.
✓ 持久化与权限
The skill is not set to always:true and is not user-invocable; autonomous invocation is allowed (platform default). It does not request persistent installation or to modify other skills. No elevated persistence privileges are declared.
安全有层次，运行前请审查代码。
运行时依赖
无特殊依赖
版本
latestv1.2.02026/3/23
降级告知+长对话恢复策略
● Pending
安装命令 点击复制
官方npx clawhub@latest install clawbrain-pro-retry
镜像加速npx clawhub@latest install clawbrain-pro-retry --registry https://cn.clawhub-mirror.com
技能文档
工具调用出错了怎么办？这个 skill 帮你的 AI 学会自己想办法。
为什么需要
我们测了 205 个场景，发现大多数模型遇到错误就放弃：
| 模型 | 出错后能自救的比例 |
|---|---|
| ClawBrain Auto | 100% |
| ClawBrain Pro | 100% |
| 单模型 A | 80% |
| 单模型 B | 76% |
| 单模型 C | 56% |
v1.0 恢复策略
第一级：智能重试
文件找不到 → 用 ls 命令查找正确文件名，再重试命令不存在 → 试替代命令（python→python3, pip→pip3）
缺少依赖 → 先安装再执行
权限不足 → 不硬来，告诉你需要什么权限
第二级：循环检测
- bigram Jaccard 相似度检测重复回复
- 连续 3 次工具调用失败，自动注入恢复指令
- 10 次循环警告、20 次严重警告、30 次熔断
第三级：双模型共识
连续失败 → 换另一个模型帮忙（ClawBrain 的多模型 fallback 链）fallback 顺序：自动选择可用的备选模型
模型健康熔断（v1.0 新增）
- 滑动窗口监控后端模型成功率和响应时间
- 5 次连续失败 → 自动熔断 60 秒
- 半开探测 → 成功则恢复，失败则继续熔断
GET /v1/health/backends查看实时健康状态
降级透明告知（v1.2 新增）
- 当后端模型或记忆服务降级时，明确告知用户当前状态
- 说明降级对当前任务的影响和预计恢复时间
- 不静默失败，不让用户困惑
长对话截断恢复（v1.2 新增）
- 超长对话被截断时，自动从知识图谱恢复关键上下文信息
- 确保截断后的对话不丢失重要背景和决策历史
- 恢复后告知用户哪些信息已从记忆中补回
核心原则
- 绝不重复同样的错误
- 每次出错最多试 2 种不同方案
- 检测到循环立即跳出
- 实在搞不定就说清楚为什么
全自动体验
这个 skill 提供错误恢复的指导。想要全自动的错误恢复（AI 自己搞定，不用你管），接入 ClawBrain 的编排引擎：
- 自动检测错误类型和模式
- 自动选择恢复策略
- 循环检测和熔断保护
- 搞不定会请另一个模型帮忙
https://clawbrain.dev/dashboard
数据来源：ClawHub ↗ · 中文优化：龙虾技能库
OpenClaw 技能定制 / 插件定制 / 私有工作流定制
免费技能或插件可能存在安全风险，如需更匹配、更安全的方案，建议联系付费定制