Minos.NET.OpenAI 1.2.0

Minos

Named after Minos, the judge of the dead in Greek myth.

Unofficial .NET client for decision models: send a state and typed questions (Noul, Choice, Score) and get calibrated, typed answers back.

Minos speaks the /v1/systemone protocol of TypeSafe's Jev: to TypeSafe directly, through OpenRouter, or to a server you run, such as a vLLM Semantic Router. The core is provider-neutral, and Minos.NET.SystemOne is the client for this protocol. Minos.NET.OpenAI is the client for OpenAI's Decisions API, /v1/decisions, which is in public beta; its provider name in logs and telemetry is openai.

You declare the questions as a C# type, and a source generator writes the request and the answer parsing at compile time, so generated question sets need no reflection and the client runs under Native AOT.

Not affiliated with TypeSafe AI. Minos is a community project. TypeSafe publishes official SDKs for Python and JavaScript; see docs.typesafe.ai.

Status: version 1.0. The public API is stable and follows semantic versioning. Coming from 0.x? See Upgrading to 1.0.

Packages

What you get

• Typed questions declared as a C# type, with a source generator.
• One provider-neutral IDecisionClient, with presets for TypeSafe, OpenRouter and local servers.
• A pipeline with retries, logging, OpenTelemetry traces and metrics, caching and rate limiting.
• Dependency-injection registration, default and keyed.
• The MIN analyzers and their code fixes.
• Native AOT, with allocation budgets enforced in CI.
• A conformance suite, shipped as minos conformance in the Minos.NET.Tool dotnet tool, that checks any decision endpoint.

Install

dotnet add package Minos.NET.SystemOne
dotnet add package Minos.NET.DependencyInjection

For OpenAI's Decisions API, add Minos.NET.OpenAI instead of, or beside, Minos.NET.SystemOne. The package is built from OpenAI's Decisions guide and is not yet tested against the live API:

dotnet add package Minos.NET.OpenAI

Two optional pipeline stages ship as their own packages: Minos.NET.Caching answers repeated decisions from an in-process IMemoryCache and an optional IDistributedCache, such as Redis, and Minos.NET.RateLimiting limits the call rate with a RateLimiter.

You need the .NET 10 SDK, and any IDE or build that hosts the [Questions] generator must host Roslyn 5.0 or later; Getting started has the details.

Example

using Minos;
using Minos.SystemOne;

public enum Lane
{
    [Criterion("Payments, invoices and refunds")]
    Billing,

    [Criterion("Bugs, outages and integrations")]
    Technical,
}

[Questions]
public partial record Triage
{
    [Noul("Does this convey urgency?")]
    public partial Noul IsUrgent { get; }

    [Choice("Which queue should handle this?")]
    public partial Choice<Lane> Queue { get; }
}

public static class ReadmeExample
{
    // A one-off run. The client reads TYPESAFE_API_KEY. A long-running app creates one client and shares it, or
    // registers it with AddSystemOneDecisionClient.
    public static async Task<string> RunAsync(string message, CancellationToken cancellationToken)
    {
        using var client = new SystemOneDecisionClient();
        return await RouteAsync(client, message, cancellationToken);
    }

    public static async Task<string> RouteAsync(IDecisionClient client, string message, CancellationToken cancellationToken)
    {
        var result = await client.EvaluateAsync<Triage>(message, cancellationToken);
        return result.IsFailure
            ? $"{result.Error.Kind}: {result.Error.Message}"
            : $"urgent: {result.Value.IsUrgent.Value}, queue: {result.Value.Queue.Value}";
    }
}

Documentation

The guide lives at marcelroozekrans.github.io/Minos.NET:

• Getting started: install Minos.NET.SystemOne, point it at TypeSafe, OpenRouter or a server of your own, and run your first typed evaluation.
• Question types: Noul, Choice and Score, what each answer holds, and how confidence differs from probability.
• Typed evaluation: declare questions as a C# type, give them a typed state, and pick the EvaluateAsync overload.
• Question sets at run time: build a question set from data with the builder, evaluate it, and read answers through handles.
• The client and its errors: create and configure a SystemOneDecisionClient or an OpenAIDecisionClient, retries and time-outs, every kind of DecisionError, and the raw request API.
• Providers: the presets, servers you run and OpenAI, the capability matrix, images, and switching providers by configuration.
• Labelled datasets: the JSONL format for items with expected answers, the members of a line, how the loader reports problems, and the format version.
• Dependency injection: register IDecisionClient in a .NET host, key several clients, and bind options from configuration.
• The client pipeline: the stages every call goes through, building your own pipeline, writing a stage, and upgrading to 1.0 and from 0.7.
• Logging, traces and metrics: what the client logs, which spans and metrics it emits, and what it never records.
• Native AOT and allocations: what Native AOT compatibility means, the one reflection the client uses, and the allocation budgets that guard it.
• Diagnostics: every MIN analyzer rule with its severity, the provider rules the client checks at run time, the two code fixes, and how to suppress a rule.
• Testing your code: test code that calls Jev with a fake IDecisionClient or a real SystemOneDecisionClient over a canned HTTP reply.
• Patterns: four ways to use Jev answers, each with a guide of its own.
• Speculative fan-out: ask every question you might need in one request and read only the answers that matter.
• Confidence routing: gate each action on the answer confidence, with a threshold sized to the cost of being wrong.
• Composite scoring: break a judgement into small Scores and combine them with weights you own.
• Intent routing: make a cheap first decision that sends each request to code, a model or a person.
• Samples: four runnable cookbook samples that replay recorded Jev answers offline, one of them against three providers chosen by configuration.
• Performance: what the benchmarks measure, how to run them, and what the client costs per call.
• Conformance: the suite that checks any decision endpoint's shapes, probabilities, errors and models, and how to run it against your own.

Samples

Runnable cookbook apps live in samples/; all four replay recorded answers, so they run offline, and one of them runs against three providers by configuration.

Building

Building this repository needs the SDK version pinned in global.json (10.0.401 or later, via rollForward: latestMinor). The [Questions] generator compiles against Roslyn 5.0, so builds and IDEs must host Roslyn 5.0 or later — Visual Studio 2026 version 18.0 is the first release that qualifies. An older IDE shows generator errors even when dotnet build succeeds.

Maintainers: RELEASING.md describes how a release is published.

Testing

dotnet test runs the unit tests, the generator tests, the analyzer and code-fix tests, the WireMock integration tests, and the pack tests, which pack the library and check the package's contents. A solution-wide dotnet test never makes a billed call: the live smoke tests in tests/Minos.NET.Live.Tests call the real APIs and only run when MINOS_LIVE=1 is set and the provider's key (TYPESAFE_API_KEY or OPENROUTER_API_KEY) is set; otherwise every one of them reports skipped. To run them locally, opt in explicitly with the key set: MINOS_LIVE=1 dotnet test tests/Minos.NET.Live.Tests. Each run makes a few small billed evaluations. Override the model with MINOS_LIVE_MODEL, or MINOS_LIVE_OPENROUTER_MODEL for OpenRouter. The Luna test through OpenRouter also needs MINOS_LIVE_LUNA=1, because Luna is unavailable to accounts that allow only zero-data-retention endpoints. There is no live OpenAI test yet. Maintainers can also run them in CI with the manual Live smoke workflow, which reads the keys from the live-api environment; restrict that environment's deployment branches to main.

About the name

In Greek mythology, Minos was king of Crete and, after his death, one of the three judges of the underworld. Every soul came before him and left with a verdict. That is what this library does. You hand it a state and a few typed questions, and a decision model returns a verdict for each one: yes or no, which option, how severe, with a calibrated confidence.

Minos does not care which model judges. TypeSafe's Jev, OpenAI's Decisions API or a model running on your own machine all answer the same typed questions. The name also keeps it in the same Cretan story as its siblings: Minos had Daedalus build the Labyrinth, and set Talos to guard his island.

License

MIT