Native reasoning controls and history

CompletionConfig accepts six distinct ReasoningEffort values. Lowercase string values also work. Leaving reasoning_effort=None sends no reasoning control; this does not mean reasoning is disabled or imply equivalent provider defaults.

Effort	Ollama	oMLX
Disabled	think: false	enable_thinking: false
Enabled	think: true	enable_thinking: true
Low	think: "low"	reasoning_effort: "low"
Medium	think: "medium"	reasoning_effort: "medium"
High	think: "high"	reasoning_effort: "high"
XHigh	think: "xhigh"	reasoning_effort: "xhigh"

These controls apply to ordinary, streaming, recovery and structured output. Structured JSON remains in content/object; reasoning stays separate.

Explicit model profiles

from mojentic.llm import CompletionConfig, ReasoningEffort, ReasoningProfile

config = CompletionConfig(
    reasoning_profile=ReasoningProfile.qwen3_8(),
    reasoning_effort=ReasoningEffort.XHigh,
    retain_reasoning=True,
)

Qwen3.8 supports exactly Disabled, Low, Medium and XHigh, plus native history. Enabled and High fail configuration validation for this profile. Omitted effort remains unspecified. Profiles are chosen by the caller, never inferred from a model identifier or alias. A custom ReasoningProfile declares name, supported_efforts and supports_reasoning_history. Unsupported efforts or retention fail before HTTP. Without a profile the gateway forwards exact controls and the provider decides model support.

Native history and tool continuation

Gateway responses expose native reasoning in LLMGatewayResponse.thinking. Assistant history uses LLMMessage.thinking; visible text stays in content, with tool calls and results in their own fields. Ollama serializes assistant reasoning as thinking; oMLX serializes it as reasoning_content.

Broker retention is off by default. Pass retain_reasoning=True to generate or generate_stream to attach completed tool-turn reasoning to assistant history before the next request. Streamed visible output remains text only. Partial, failed, capped or cancelled turns never supply retained reasoning. Native history is never reconstructed from visible content. For an outer turn managed by your application, explicitly copy a completed gateway response's thinking into an assistant message when desired.

With retention enabled oMLX requests chat_template_kwargs: {"preserve_thinking": true}. This records the client request; forced server settings can override it. Tool identifiers and arguments remain distinct from reasoning. oMLX carries tool-call identifiers; the Ollama SDK's tool protocol uses function names and arguments without native call IDs.

Migration from the old effort switch

Earlier Python versions mapped every Ollama level to think: true. Use ReasoningEffort.Enabled for that boolean behavior. Named levels now send their exact strings. Qwen3.8 callers should choose its explicit profile and a supported level rather than Enabled. Ollama ordinary tool arguments now retain JSON types, and assistant history includes every tool call in its original order.

OpenAI retains existing Low/Medium/High handling and model parameter adaptation. It rejects Disabled, Enabled, XHigh, explicit profiles, retention and native assistant history with ValueError; it does not accept the oMLX extension. Existing temperature/token defaults and deprecated individual keyword arguments are unchanged by this reasoning correction. Metadata/request fidelity is separate 3.0.0 work.

Local tests execute native-reasoning-v1 through public gateway HTTP and broker continuations. External consumers and live model servers have not verified this set. See CONFORMANCE-3.0.0.md for identities, commands and remaining verification; the coordinated publication hold remains.