dsh-llm-capabilities

When wiring up a self-hosted gateway in llm-pi-ai, the official settings page only exposes contextWindow for editing. Two per-model fields that actually drive routing quality — reasoning levels (thinking intensity) and vision input modality — are buried in providers…models[] with no UI affordance. The practical result: the intensity slider stays greyed out with a "no selectable thinking levels" note, and a model that clearly supports image input is still marked text-only locally, so every image attachment is rejected pre-flight with UNSUPPORTED_CONTENT before it ever reaches the endpoint.

dsh-llm-capabilities adds a single panel under Settings → Model Capabilities. It fetches the gateway's /models endpoint server-side (credential resolved host-side, the secret never leaves the host), merges the result with pi-ai's own catalog knowledge as a complementary source, and lets you toggle per-model: pick from off/low/medium/high/xhigh/max reasoning levels with inline wire-spelling edits, set a three-state vision mode (inherit / image / text-only), and adjust context length and max tokens. Models listed by the endpoint but not yet configured appear with a "not configured" badge; a single Apply writes everything into llm-pi-ai's providers…models[] and the intensity slider and image admission update immediately.

Built for developers running self-hosted OpenAI-compatible or DeepSeek gateways on llm-pi-ai who need the routing layer to genuinely respect thinking intensity and vision capability — without hand-editing raw JSON.