dsh-provider-hub

Every time a DSH user adds a new LLM gateway, they have to hand-edit llm-pi-ai.providers in settings.yaml, filling in protocol, endpoint, model list, and reasoning tiers, and a misconfiguration means a restart. With multiple gateways that YAML quickly turns into an entangled block of text. dsh-provider-hub lifts the entire workflow into the Provider Hub panel on the left sidebar: a card-style console where you add gateways one by one, expand a segmented editor covering basic fields, request headers, model table, and raw JSON, and save to hot-swap routes, no YAML required.

The panel ships a built-in catalog spanning GLM, Claude, GPT, Qwen, DeepSeek, Kimi, Gemini, and more. Typing a model name triggers prefix-aware autocomplete that carries contextWindow, maxTokens, input modality, and reasoning-effort parameters; a one-click discovery call fetches the full upstream /models list, or you can hand-write every parameter in the JSON view, with list and JSON staying bidirectionally in sync. Each gateway independently configures protocol, Base URL, a fully custom User-Agent, API key, and extra headers. For gateways that whitelist by UA, the plugin sends the exact UA you set on that gateway, bypassing the default DSH attribution header. The adapter classifies every upstream failure into a canonical code; transient errors are automatically retried with exponential backoff and jitter, while deterministic failures are reported as-is without wasting turns.

Built for DSH users who run two or more LLM gateways simultaneously, need per-gateway UA or header differentiation, and would rather not keep wrestling with provider YAML.