Release Notes: v0.18.0¶

Release Date: 2026-07-05

Highlights¶

Reasoning & Extended Thinking: Unified ReasoningEffort and ThinkingConfig fields for controlling model reasoning depth across all providers

New Features¶

Reasoning Effort¶

Control how deeply models reason about problems using the unified ReasoningEffort field:

import omnillm "github.com/plexusone/omnillm-core"

effort := omnillm.ReasoningEffortHigh
response, err := client.CreateChatCompletion(ctx, &omnillm.ChatCompletionRequest{
    Model:           "o1-preview", // or any reasoning-capable model
    ReasoningEffort: &effort,
    Messages:        messages,
})

Supported Values¶

Constant	Value	Use Case
`ReasoningEffortNone`	`"none"`	Disable reasoning, fastest responses
`ReasoningEffortLow`	`"low"`	Light reasoning, quick analysis
`ReasoningEffortMedium`	`"medium"`	Balanced reasoning (default for most tasks)
`ReasoningEffortHigh`	`"high"`	Deep reasoning for complex problems

Extended Thinking (Anthropic)¶

For fine-grained control over Anthropic's extended thinking feature:

budget := int64(10000)
response, err := client.CreateChatCompletion(ctx, &omnillm.ChatCompletionRequest{
    Model: "claude-sonnet-4-6",
    Thinking: &omnillm.ThinkingConfig{
        Type:         omnillm.ThinkingTypeEnabled,
        BudgetTokens: &budget, // Max tokens for thinking
    },
    Messages: messages,
})

Thinking Types¶

Constant	Behavior
`ThinkingTypeEnabled`	Always use extended thinking with specified budget
`ThinkingTypeDisabled`	Never use extended thinking
`ThinkingTypeAdaptive`	Let the model decide when thinking helps

Provider Support Matrix¶

Provider	ReasoningEffort	ThinkingConfig	Notes
OpenAI	Yes	-	Maps directly to `reasoning_effort` parameter
Anthropic	Yes	Yes	Maps to extended thinking; ThinkingConfig preferred
X.AI (Grok)	Yes	-	Native `reasoning_effort` support
GLM (Zhipu)	Yes	-	Native `reasoning_effort` support
Kimi (Moonshot)	Yes	-	Native `reasoning_effort` support
Qwen (Alibaba)	Yes	-	Native `reasoning_effort` support
Ollama	Yes	-	Passed through for compatible models

Cross-Provider Mapping¶

When using ReasoningEffort with Anthropic (which natively uses ThinkingConfig), the values are automatically mapped:

ReasoningEffort	Anthropic Thinking
`none`	Disabled
`low`	Adaptive
`medium`	Adaptive
`high`	Enabled (8192 token budget)

Usage Examples¶

Basic Reasoning Control¶

// High reasoning for complex math
effort := omnillm.ReasoningEffortHigh
resp, _ := client.CreateChatCompletion(ctx, &omnillm.ChatCompletionRequest{
    Model:           "o1-preview",
    ReasoningEffort: &effort,
    Messages: []omnillm.Message{
        {Role: omnillm.RoleUser, Content: "Prove that sqrt(2) is irrational"},
    },
})

Anthropic Extended Thinking¶

// Explicit thinking budget
budget := int64(16000)
resp, _ := client.CreateChatCompletion(ctx, &omnillm.ChatCompletionRequest{
    Model: "claude-sonnet-4-6",
    Thinking: &omnillm.ThinkingConfig{
        Type:         omnillm.ThinkingTypeEnabled,
        BudgetTokens: &budget,
    },
    Messages: messages,
})

Adaptive Thinking¶

// Let the model decide
resp, _ := client.CreateChatCompletion(ctx, &omnillm.ChatCompletionRequest{
    Model: "claude-sonnet-4-6",
    Thinking: &omnillm.ThinkingConfig{
        Type: omnillm.ThinkingTypeAdaptive,
    },
    Messages: messages,
})

Installation¶

go get github.com/plexusone/omnillm-core@v0.18.0

Documentation¶

Upgrade Notes¶

This is a backwards-compatible release. No breaking changes. The new ReasoningEffort and Thinking fields are optional and do not affect existing code.