跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

retry-budget-policy

Design and review retry budgets for agent steps, tools, providers, and transport calls. Use when classifying retryable errors, enforcing max attempts, respecting Retry-After, applying backoff and jitter, handling timeouts, or preventing retries for permission, validation, business, or user-cancelled failures.

HsienW · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add HsienW/ai-agent-engineering-playbook --skill retry-budget-policy

技能说明


name: retry-budget-policy description: Design and review retry budgets for agent steps, tools, providers, and transport calls. Use when classifying retryable errors, enforcing max attempts, respecting Retry-After, applying backoff and jitter, handling timeouts, or preventing retries for permission, validation, business, or user-cancelled failures.

Retry Budget Policy

Skill Interface

  • Name: retry-budget-policy.
  • Description: Design and review retry budgets for agent steps, tools, providers, and transport calls when classifying retryable errors, enforcing max attempts, respecting Retry-After, applying backoff and jitter, handling timeouts, or blocking retries for permission, validation, business, and user-cancelled failures.
  • Parameters: Error taxonomy, retryable codes, max attempts, max elapsed time, backoff strategy, jitter policy, Retry-After handling, cancellation signal, persisted retry state, idempotency requirements, and verification cases.
  • Instructions: Use this skill before adding retry behavior. Classify the failure first, retry only allowed classes, stop at budget limits, carry cancellation through waits and upstream calls, persist retry state when needed, and test retryable, non-retryable, exhausted, cancelled, and idempotent paths.

Retries are a budgeted recovery mechanism. Each policy needs an error taxonomy, a budget, a backoff strategy, and a stop rule.

Error Classification

Classify failures before retrying:

Error classRetryStrategy
TimeoutYesExponential backoff with jitter
Rate limitedYesRespect server-provided retry hints
Temporary upstream failureYesLimited retry budget
Structured output parse errorConditionalOne repair or retry attempt
Structured output validation errorConditionalRepair with explicit hint
Permission deniedNoEscalate or terminate
Business rejectedNoReturn explainable result
User cancelledNoStop and compensate if needed
Invalid inputNoAsk for correction or return validation error

Policy Shape

type RetryPolicy = {
  maxAttempts: number;
  maxElapsedMs: number;
  retryableCodes: string[];
  backoffStrategy: 'exponential' | 'fixed' | 'retry_after';
  jitter: boolean;
};

Track attempts, elapsed time, last error, and next retry time per step or operation.

Rules

  • Retry only errors classified as retryable.
  • Hard-stop when maxAttempts or maxElapsedMs is exceeded.
  • Carry cancellation signals through sleeps and upstream calls.
  • Do not retry permission, policy, business rejection, or user cancellation.
  • Persist retry state when a retry can span process restarts.
  • Emit structured events for retry scheduled, retry started, retry succeeded, and budget exhausted.
  • Make retries idempotent before retrying side effects.

Verification

Test timeout retry, rate-limit retry hints, temporary failure recovery, non-retryable errors, budget exhaustion, cancellation during delay, jitter boundaries, persisted retry state, and interaction with idempotency.