Universal.Operative.Sdk.Discrete.MoonshotAI 1.0.0

About

Universal.Operative.Sdk.MoonshotAI plugs Moonshot AI's Kimi models into Universal.Operative.Sdk via an IModel adapter (Model) backed by the Universal.MoonshotAI.Client transport.

Unlike OpenAI, Moonshot has no Responses-equivalent endpoint — Chat Completions is the only surface — so there's a single Model here rather than a Chat/Responses pair.

How to Use

Installation

dotnet add package Universal.Operative.Sdk.MoonshotAI

Configuration

using Universal.Operative.Sdk;
using Universal.Operative.Sdk.MoonshotAI;

IModel model = new Model(apiKey, new Model.Options
{
    ModelId = "kimi-k3",
    MaxCompletionTokens = 4096,
});

The constructor also accepts an existing Universal.MoonshotAI.Client.V1.V1Client in place of an API key, so you can reuse a pre-built client (custom endpoint — e.g. api.moonshot.cn — headers, etc.) across models:

using Universal.MoonshotAI.Client.V1;

V1Client client = new V1Client("https", "api.moonshot.cn", apiKey);
IModel model = new Model(client, new Model.Options { ModelId = "kimi-k3" });

Usage

Drop the model into EngineBuilder like any other IModel:

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

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

Reasoning models

kimi-k3 takes a ReasoningEffort hint ("low"/"high"/"max"); kimi-k2.6/kimi-k2.5 take a Thinking toggle instead (kimi-k2.7-code always reasons and exposes neither). Either way, the chain-of-thought comes back as a ReasoningBlock in the settled Message, same as any other provider's extended thinking:

IModel reasoningModel = new Model(apiKey, new Model.Options
{
    ModelId = "kimi-k2.6",
    Thinking = new ThinkingConfiguration { Type = ThinkingTypes.Enabled },
});

Structured output

Model.Options.ResponseFormat is the escape hatch for JSON object output, forwarded verbatim as response_format. Moonshot's own JSON-mode guide only documents {"type": "json_object"} — describe the desired shape in the prompt itself, since this parameter alone doesn't communicate a schema to the model.

Known gaps

Universal.MoonshotAI.Client has no typed content-block classes (see its own Package.md), so this adapter builds image_url content parts by hand from ImageBlock; DocumentBlock, AudioBlock, and VideoBlock throw NotSupportedException rather than guessing at an unverified wire shape. Moonshot's built-in $web_search tool isn't wired up here — as of this writing it reliably fails server-side even when built exactly to Moonshot's documented protocol (confirmed via a raw request bypassing the client library entirely); wire it up once that's fixed upstream.

License

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