retry-optimizer

--- name: retry-optimizer description: Optimize retry patterns for flaky or slow agents/services. Use when you see fixed-interval retries, repeated 30s/60s wait loops, Retrying... messages, wrapper timeouts, or performance bottlenecks caused by retry storms. --- # Retry Optimizer Optimize brittle retry loops into efficient backoff with jitter, circuit breaking, and bounded retries. ## When to Use - Wrapper/agent keeps printing `Retrying in 30s...` - Same command fails repeatedly with fixed sleep - Retry storms under load - No timeout / max-retry guard ## Quick Start ```bash # Analyze current retry config node skills/retry-optimizer/scripts/optimize.js analyze <config.json> # Generate optimized config node skills/retry-optimizer/scripts/optimize.js generate <config.json> # Validate a retry config node skills/retry-optimizer/scripts/optimize.js validate <config.json> ``` ## Config Format ```json { "retries": 3, "initialDelayMs": 1000, "maxDelayMs": 30000, "backoffFactor": 2, "jitterMs": 500, "timeoutMs": 5000, "retryOn": ["ECONNRESET", "ETIMEDOUT", "5xx"], "abortOn": ["4xx", "auth_failed"] } ``` ## Strategy 1. Replace fixed delay with exponential backoff 2. Add jitter to prevent synchronized retries 3. Cap max delay to avoid excessive waits 4. Set max retry limit to prevent infinite loops 5. Add per-attempt timeout 6. Classify errors: retryable vs abort ## Benchmark Results | Scenario | Fixed Interval | Backoff + Jitter | Improvement | |----------|---------------|------------------|-------------| | 30s fixed (5 attempts) | 150,000ms | 32,208ms | **78.5% faster** | | 60s fixed (5 attempts) | 300,000ms | 32,580ms | **89.1% faster** | | 10s fixed (5 attempts) | 50,000ms | 16,315ms | **67.4% faster** | Run benchmark: ```bash node skills/retry-optimizer/scripts/benchmark.js node skills/retry-optimizer/scripts/benchmark.js --scenarios=3 --attempts=5 ``` ## Notes - Avoid unbounded retries; always set max retries - Prefer backoff + jitter over fixed sleep - Add circuit breaker for repeated failures - Benchmark shows 67-89% reduction in total wait time

存档时间线

版本索引永久保留；内容副本只保留最新 1 份，更早版本仅留索引与哈希（存档时间线的证据链不会因此断裂）。

下载存档内容副本