dsh-auto-retry

When working with DSH, a certain class of failures is particularly confusing: the relay mislabels a recoverable transient 429 as QUOTA, or the model request appears to complete normally yet produces no visible output. The official dsh-llm-retry only handles errors declared by the provider itself; once it gives up, the request is a dead end. dsh-auto-retry is purpose-built for that gap; it does not replace the official retry, but steps in only after the provider policy has ended and clear evidence of recoverability exists, catching transient errors with bounded extra retries.

The plugin caps extra retries (max 8), total wait time (15 minutes), and auto-continuation per turn (max 2) within a configurable budget. For mislabeled QUOTA errors it triggers a retry only when the error detail contains signatures such as rate_limit_exceeded, rate limit, or allocated quota exceeded; genuine balance exhaustion, AUTH failures, and user cancellation are never retried. Backoff prefers the provider Retry-After header, falling back to jittered exponential backoff between 1.5 and 60 seconds. When a turn ends with no visible text or hits a max-tokens truncation, it auto-continues without disguising the continuation as a human message.

It suits DSH Web profile users who regularly hit relay-side transient rate-limit mislabels or silent empty responses. The plugin does not switch models or providers, does not mutate request content, and its events integrate cleanly with existing session replay and retry-status UIs. Only one plugin should own the global request-error retry budget per profile.