v1.1.0

• Switch the provider from the openai-completions adapter to anthropic-messages (authHeader: true), the gateway's native endpoint.
• Map reasoning onto Anthropic adaptive thinking: pi thinking levels translate to Umans effort levels (pi xhigh reaches Umans max), models that can't disable thinking keep it on, and unsigned thinking blocks are preserved instead of downgraded to text.
• Clamp max output tokens below max_completion_tokens so requests never trip the gateway's hard-cap 400.
• Register gateway-handoff vision models (supports_vision: "via-handoff", e.g. umans-glm-5.2) with image input; the gateway runs the handoff server-side. Previously these were text-only. If your media proxy skips models that already report image support, it will now leave these images to the gateway.
• Skip deprecated models from the catalog.
• Add selfcheck.ts (npm test) covering the registration logic.