dsh-better-retry

In DeepSeek Harness, some upstream rate-limit responses are tagged as PI_AI_ERROR instead of RATE_LIMIT, so the built-in exponential back-off and retry pipeline never kicks in and users must resend requests by hand. dsh-better-retry reclassifies messages that match configured patterns into RATE_LIMIT, letting the existing DSH retry strategy take over seamlessly. The original request text and diagnostic context stay untouched, while retry count, back-off, and jitter remain governed by the provider settings.

The plugin offers two pattern-matching modes — full normalized match and contiguous-fragment match — and exposes two management entry points: the Settings page and an inline panel under each error row. You can add, disable, or delete patterns, and test whether a message hits an enabled rule using the same matcher that runs in the live stream. Two built-in safety rules are enabled by default to cover common rate-limit wording, and the plugin automatically rejects patterns containing authentication-failure, balance-exceeded, or cancellation semantics that should never be retried.

Best for developers running DSH Web locally or self-hosted who frequently hit upstream rate limits without automatic retry. It does not replace DSH retry executors; it only corrects the error-classification layer, making it a low-intrusion, long-running add-on for existing workflows.