Universal.Operative.Sdk.Discrete.Google.Gemini 3.0.0

About

Universal.Operative.Sdk.Discrete.Google.Gemini plugs Google's Gemini models into Universal.Operative.Sdk via an IModel adapter (Model) backed by the Universal.Google.Gemini.Client transport (Gemini's v1beta REST/SSE API).

How to Use

Installation

dotnet add package Universal.Operative.Sdk.Discrete.Google.Gemini

Configuration

using Universal.Operative.Sdk;
using Universal.Operative.Sdk.Discrete.Google.Gemini;

IModel model = new Model(apiKey, new Model.Options
{
    ModelId = Models.Gemini35Flash,
    MaxOutputTokens = 4096,
});

The constructor also accepts an existing Universal.Google.Gemini.Client.V1Beta.GeminiClient in place of an API key, so you can reuse a pre-built client (custom headers, handlers, ...) across models:

using Universal.Google.Gemini.Client.V1Beta;

GeminiClient client = new GeminiClient(apiKey);
IModel model = new Model(client, new Model.Options { ModelId = Models.Gemini35Flash });

Usage

Drop the model into EngineBuilder like any other IModel:

IEngine engine = EngineBuilder.FromDefaults(model)
    .AddBaselineToolSet()
    .Build();

await foreach (EngineEvent e in engine.ExecuteAsync(conversation))
{
    // handle events as usual — see Universal.Operative.Sdk's Package.md
}

Extended thinking

Model.Options.ThinkingConfig forwards directly to Gemini's thinkingConfig (thinking budget, whether to surface thought summaries). Thinking text round-trips through the universal ReasoningBlock.Text, the same slot Anthropic's extended thinking uses.

Gemini's thoughtSignature, however, does not live on the thought itself — verified against live responses, it's attached to whichever block concludes the thinking segment (the final answer TextBlock, or the ToolCallBlock for the function call that followed). Since neither has a native signature slot, it round-trips via the Gemini extension property added onto Block:

using Universal.Operative.Sdk.Discrete.Google.Gemini;

string? signature = someBlock.Gemini.ThoughtSignature;

You only need to read this if you're inspecting history yourself — the adapter re-attaches it to the right outbound part automatically when you replay a conversation, so tool calls and answers produced under ThinkingConfig continue to validate correctly on the next turn.

Usage accounting

Gemini reports thinking tokens (thoughtsTokenCount) and tool-use prompt tokens (toolUsePromptTokenCount) as additive counters alongside promptTokenCount/candidatesTokenCount — unlike Usage.InputTokens/OutputTokens, which mirror Gemini's own headline counters as-is. Use Usage.Gemini to read the breakdown, and GeminiUsageExtensions.ProjectCompactionTokens as a CompactionTransformer.Options.ProjectedTokensCalculator so auto-compaction accounts for thinking-token cost correctly.

Prompt caching

Gemini 2.5+ already caches long context automatically — no configuration needed; savings show up in Usage.Gemini's cachedContentTokenCount. For a guaranteed cache hit instead, create a CachedContent via GeminiClient.CreateCachedContentAsync and reference it:

IModel model = new Model(client, new Model.Options
{
    ModelId = Models.Gemini35Flash,
    CachedContentName = cachedContent.Name,
});

A referenced cache carries its own system instruction/tools, so the request omits Message.Tools/Conversation.System while it's set.

License

Free for noncommercial use under the Universal.Operative Noncommercial License. Source is closed; commercial use requires a separate license from Andrew Ong.