dsh-poor-router

Free LLM capacity is scattered across NVIDIA bonus credits, Qwen giveaway grants, GLM flash tiers, and OpenRouter :free routes, each one expiring on its own schedule. You are mid-sentence when a model flatlines; you scramble to the next provider and find it barely breathing. dsh-poor-router treats every piece of free or cheap capacity as one pooled ledger: it meters every request, samples live health per entry, and swaps the model the instant a provider dies, so your session never notices a hiccup.

Routing is driven by Thompson sampling: each model entry carries a Beta posterior built from success and fail counts, and the per-request draw is further weighted by a TTFT latency penalty and the current-hour congestion bucket, so probability—not guesswork—picks the healthiest entry at the moment of routing. Auxiliary and trivial sub-requests are sent to whichever executor model still has the thickest remaining grant, with an epsilon-explore slot every sixth call to probe unused tiers. Only when every free candidate is dead does the escape hatch release the cheapest capped paid model, and a single SMS alerts you that real money was spent. At the provider level, AUTH or QUOTA-class failures cool the endpoint for ten minutes, 429 responses for ninety seconds, and all cooldowns persist across restarts.

The plugin is stateless per request: every call starts from your configured model and may be swapped for that one invocation only—your settings are never permanently altered. It also works fully headless; the two registered agent tools function without the web panel or the input-box badge. It is built for solo developers and small teams who survive on fragmented free and budget LLM quotas. If you can afford API bills, walk away; if you live on free credits, this is your routing brain.