PineappleTwilight/dsh-llm-retry-infinite

README 摘要

dsh-llm-retry-infinite A DeepSeek Harness plugin that replaces the built-in LLM retry behavior with infinite exponential retries . Every failed LLM request is retried indefinitely with an exponential backoff that caps each individual wait at 10 minutes . Why this exists The built-in @deepseek-ai/dsh-llm-retry defaults to mode: 'normal' with a hard cap of 2 retries . For environments with transient rate limits, flaky connectivity, or provider instability, you may want the harness to keep trying until the request succeeds — without a retry ceiling. This plugin takes over the entire retry chain and never gives up. How it works 1. Intercepts every agent/request-error event from the DSH agent loop. 2. Computes an exponential backoff: min(initialDelayMs × 2^retry, 600 000) . 3. Waits the computed delay (symmetric jitter, cancellable on abort or session close). 4. Returns { kind: 'retry' } to re-attempt the request. 5. Repeats forever until success, cancellation, or plugin disposal. The plugin does not delegate to the built-in retry handler — it fully replaces it. Backoff schedule (default config) Retry Delay Approx. 1 1000 ms 1 s 2 2000 ms 2 s 3 4000 ms 4 s 4 8000 ms 8 s 5 16 000 ms 16 s…

分类