dsh-llm-failover

Running multiple LLM providers in parallel brings a silent pain point: a sporadic 429 is tolerable, but a quota-exhausted backend that keeps dying makes your request pipeline stall, and you are left guessing which provider is still alive. dsh-llm-failover turns that into a transparent background action. When a provider hits the configurable threshold of consecutive rate-limit or quota errors, the plugin cools it down, promotes the next provider in your list, and drops a brief auto-dismissing notice above the input field so you never have to hunt through logs.

Three capabilities do the heavy lifting. First, per-provider model mapping lets each fallback speak a different model, so your primary path can use deepseek-chat while the backup quietly talks deepseek-reasoner. Second, a cooldown window (cooldownMs) keeps a freshly-failed provider out of rotation until it recovers, preventing a tight retry loop against an already-saturated endpoint. Third, the last entry in the provider list is a permanent catch-all that is never switched away from, guaranteeing the system always converges and never spin-loops.

If you are already juggling two or three API keys or relay services inside DeepSeek Harness and want rate-limit and quota hiccups to stop breaking your flow, this plugin wires the failover logic into the official agent waterfall with zero extra moving parts. Configure it once in the settings card, then keep coding while the machinery handles the rest.