dsh-llm-extra-retry

In the DSH model-inference failure-recovery waterfall, upstream error responses are sometimes unrecognizable by dsh-llm-pi-ai and fall into the PI_AI_ERROR fallback code. The default policy in @deepseek-ai/dsh-llm-retry (5 bounded exponential backoff from 0.5s to 10s with 10% jitter) only acts on known error codes, so PI_AI_ERROR is passed straight through and the request fails with no additional retry opportunity.

dsh-llm-extra-retry injects a fully independent recovery policy into the waterfall: upon detecting PI_AI_ERROR it waits 20 seconds and retries the request once; every other error code is passed through via next() untouched, preserving the default strategy behavior exactly. The two policies carry separate policyKeys, so their bookkeeping never interferes with each other. Aborting the request or uninstalling the plugin immediately cancels any pending backoff, and DELAY_MS and MAX_RETRIES are exposed as constants for fine-tuning wait duration and retry count.

This plugin is well-suited for developers whose DSH workflow frequently hits transient upstream failures that land on an unrecognized error code and slip past the default retry strategy. It adds one extra chance exactly where the default policy cannot reach, with zero intrusion and zero conflict.