[zemaj]

Ensemble's catalog and provider adapters could not select gpt-6-sol, gpt-6-luna, or claude-opus-5-5. This adds those exact IDs with current capabilities and pricing, maps the supported reasoning settings, and sends Opus 5.5 structured output through Anthropic's native schema format. Explicit modelSettings.max_tokens now reaches the OpenAI Responses API as max_output_tokens, including for existing models such as gpt-5.6-terra.

GPT-6 Sol and Luna use the documented 272K-input pricing tier, support explicit none and low effort, and normalize the older minimal budget mapping to low. Opus 5.5 uses adaptive thinking with medium default; attempts to disable thinking fail clearly. Existing Claude schema transport remains unchanged for older models.

Validation: npm run lint:fix, npm test (734 passed, 3 skipped), npm run build, and git diff --check origin/main...HEAD passed on source 85a78ebb6cf707017bb9f08b7853ab0d3c0b07fb. Live image-to-strict-JSON calls succeeded for Sol/Luna at none and low, Terra at none, and Opus 5.5 at medium; every call logged the exact provider-returned model, response/message ID, usage, and the requested 8,192-token wire limit.