dsh-retry-boost

Pairing DSH with any LLM gateway, transient 429 rate-limiting and stream disconnections are a daily reality. The real trap: providers like SenseNova label these errors as insufficient_quota, which DSH maps to the QUOTA code—a code absent from the default retryableCodes list. The turn dies on the first attempt, and every manual resend can just hit the same wall. The upstream issue remains unresolved, leaving you to babysit each failed turn.

dsh-retry-boost takes a straightforward approach: at startup it hot-injects a widened retryPolicy into every llm-pi-ai provider via the settings namespace, so DSH built-in retry finally has QUOTA in scope. Backoff is exponential (1 s initial, 60 s cap, 20 percent jitter by default), and three strategies—fill, boost, force—let you choose how aggressively the plugin overwrites existing policies. Providers added at runtime are automatically covered through the llm/adapters-updated listener, no restart needed. The plugin also scans variant routes (vision-toolkit- and similar) for a missing QUOTA code and logs a clear warning, so you always know the coverage boundary. Every injected value is visible and editable in settings.yaml, writes are idempotent, and the plugin never fights your manual configuration.

Who it is for: anyone running DSH against an LLM gateway that throws transient 429, insufficient_quota, stream-cut, or timeout errors—SenseNova, self-hosted gateways, and many relays all fall into this category. A single install command gives every current and future provider automatic exponential-backoff retry, eliminating the need to manually resend failed turns. For an extra safety net, layer dsh-chat-continue on top; the two are complementary and coexist without conflict.