dsh-cpa

When you rely on CLI Proxy API as a multi-account LLM backend behind dsh, stale model lists, hard errors on quota exhaustion, and transient management-API outages that break routing are frequent headaches. dsh-cpa folds CPA provider registration, /v1/models periodic sync, intelligent routing, and automatic failover into the dsh plugin layer, so application code no longer needs hand-rolled retry and fallback logic.

The core of the plugin is a transparent request pipeline: it builds a candidate order from model capabilities, context windows, and per-account quota snapshots already available in CPA. A candidate switch only fires when no streamed content has been produced and the error matches quota-exhausted, rate-limited, server-side, transport, or empty-response categories. Once the stream has started, replay is never attempted, protecting agents from duplicate content or side effects. Diagnostics, preflight checks, sanitized execution records, and a read-only summary endpoint keep the full chain inspectable from the composer status line.

Built for teams that manage multiple provider accounts through CPA and want dsh to handle routing resiliency and quota visibility natively. The plugin only consumes desensitized data from CPA and never duplicates its credential-management responsibilities, keeping each tool in its lane.