Shared reasoning-effort vocabulary and conservative capability policy.
The established low/medium/high/xhigh behavior is unchanged. max is an
explicit intent, never an omission: only exact gpt-6-astra through Pixir's
ChatGPT/Codex Responses backend supports it. Strict open_responses retains
its existing no-reasoning boundary, even at an official OpenAI endpoint.
Unknown models, custom providers, and legacy nonstandard routes gain no max
capability by inference. This describes local admission, not live acceptance.
Summary
Functions
Effort choices for one frozen effective Provider selection.
Effort choices without performing Config or Registry reads.
Whether a value is a known explicit effort, independently of capability.
Known intent ids; capability must still be checked against the effective selection.
The established model-independent vocabulary (legacy Config API backing).
True only for the explicitly supported Astra Responses path.
Parse a known effort; nil/default explicitly mean omission.
Validate max against the effective selection, retaining legacy non-max behavior.
Validate child runtime intent after spec/opts precedence, using the same selection as Turn.
Functions
Effort choices for one frozen effective Provider selection.
Effort choices without performing Config or Registry reads.
Whether a value is a known explicit effort, independently of capability.
Known intent ids; capability must still be checked against the effective selection.
The established model-independent vocabulary (legacy Config API backing).
True only for the explicitly supported Astra Responses path.
Parse a known effort; nil/default explicitly mean omission.
Validate max against the effective selection, retaining legacy non-max behavior.
Validate child runtime intent after spec/opts precedence, using the same selection as Turn.