v1.20.4.1

n v1.20.4, selecting gpt-6-luna or another unrecognized model name silently omitted extraction reasoning, including an explicit reasoning: {effort: none}. The provider's model default then applied.

extract.ai now defaults to reasoning effort none for every model and sends caller, saved-model, and configured reasoning without a model-name/capability gate. Responses sends reasoning; Chat Completions sends the equivalent reasoning_effort. Explicit settings retain precedence. The patch includes regression coverage, matching documentation/schema descriptions, updated live-test expectations, and the version bump to 1.20.4.1.

Full Changelog: v1.20.4...v1.20.4.1