tool-call-retry

Auto retry & fix LLM tool calls with exponential backoff, format validation, error correction, boost tool call success rate by 90%

Security Scan

The skill is an internally consistent retry-wrapper for tool calls (exponential backoff, validation, optional error-fixing) and its code, docs, and requirements align with that purpose.

Purpose & Capability

Name, description, SKILL.md, README, and the TypeScript implementation all describe and implement the same functionality: a retry wrapper around a provided tool function with validation, error-handler hooks, and an in-memory idempotency cache. There are no unrelated credentials, binaries, or config paths requested.

Instruction Scope

Runtime instructions and examples only describe calling a provided tool function (toolFn) and optional validator/errorHandler callbacks. The skill does not instruct reading unrelated files, env vars, or sending data to hidden endpoints. Custom callbacks (validatorFn/errorHandlerFn) run user-supplied code and therefore have normal scope to make network calls or access data in their closures — this is expected behavior for this wrapper but should be considered by the integrator.

Install Mechanism

No explicit install spec; this is instruction/code-only. The implementation imports zod from deno.land via an HTTPS URL (https://deno.land/x/zod@v3.22.4). Using a well-known public host like deno.land is common, but it does mean a remote fetch of that dependency will occur at runtime — verify you trust the referenced version.

Credentials

The skill requests no environment variables or credentials. That matches its purpose. Note: user-supplied callbacks (toolFn, errorHandlerFn, validatorFn) run arbitrary code and may themselves access environment variables or secrets available in the runtime; that is expected but not caused by the skill.

Persistence & Privilege

The skill keeps an in-memory idempotencyCache (Map) for idempotencyKey-based deduplication. This cache is process-local and ephemeral (no persistent storage/privileged behavior). 'always' is false and the skill does not modify other skills or system configs.

Assessment

This skill appears to do exactly what it claims: wrap a provided tool function with retry/validation and optional error-fixing. Before installing, consider the following: 1) The implementation imports zod from deno.land at runtime — confirm you trust that remote dependency and version. 2) validatorFn and errorHandlerFn are user-supplied callbacks that execute arbitrary code and therefore can access any runtime secrets or make network calls; ensure callbacks you provide are safe and don't leak sensitive data. 3) Idempotency is handled in-memory only (per process); for long-lived deduplication across restarts you’ll need your own persistent store. 4) If your toolFn has side effects (database writes, payments, emails), use idempotencyKey and tune maxRetries to avoid duplicate side effects. 5) The skill logs to console and uses setTimeout for delays — no hidden network exfiltration is present in the skill itself. Overall it is coherent and small, but review/verify any callbacks you pass in and the remote zod dependency before use.

🔥 工具调用自动重试器

核心亮点

1. ✅ 成功率提升90%+：内置指数退避重试、格式校验、错误自动修复，解决Agent工具调用不稳定的核心痛点
2. 🛡️ 全链路异常兜底：自定义错误处理、参数修复逻辑，支持复杂场景下的错误自愈
3. ⚡ 零侵入增强：无需修改原有工具代码，一行封装即可获得重试能力，性能开销<1ms
4. 🔑 幂等性保证：支持幂等性键，避免重复调用导致的副作用

🎯 适用场景

• 所有调用外部API/工具的Agent场景
• 不稳定的第三方服务调用
• 大模型工具调用格式错误自动修复
• 高可靠性要求的Agent执行链路

📝 参数说明

参数	类型	必填	默认值	说明
toolFn	Function	是	-	要执行的工具函数，返回Promise
args	any	否	{}	调用工具的参数
maxRetries	number	否	3	最大重试次数，1-10
initialDelayMs	number	否	1000	初始重试延迟，100-10000ms
validatorFn	Function	否	()=>true	结果校验函数，返回true表示结果合法
errorHandlerFn	Function	否	undefined	错误处理函数，可返回修复后的参数或中止重试
idempotencyKey	string	否	undefined	幂等性键，相同键的调用只会执行一次

💡 开箱即用示例

基础用法（零配置）

const fetchWeather = async (params: { city: string }) => {
  const res = await fetch(`https://api.weather.com/${params.city}`);
  return res.json();
};

const result = await skills.toolCallRetry({
  toolFn: fetchWeather,
  args: { city: "Beijing" }
});

带结果校验

const result = await skills.toolCallRetry({
  toolFn: callLLM,
  args: { prompt: "Generate JSON output" },
  validatorFn: (res) => typeof res === "object" && res !== null && res.code === 0,
  maxRetries: 5
});

高级用法（错误自动修复）

const result = await skills.toolCallRetry({
  toolFn: callDatabase,
  args: { sql: "SELECT * FROM users" },
  errorHandlerFn: async (error, attempt) => {
    if (error.message.includes("SQL syntax error")) {
      // 自动修复SQL语法
      const fixedSql = await fixSqlWithLLM(error.message);
      return { args: { sql: fixedSql } };
    }
    if (attempt >= 2) {
      // 重试2次失败后中止
      return { abort: true };
    }
  }
});

🔧 技术实现说明

• 采用指数退避重试算法，避免服务被打垮
• 全链路类型安全，所有参数带Zod校验
• 支持自定义校验和错误修复逻辑，可扩展性强
• 轻量无依赖，仅200行代码，无额外性能开销