AnointedAutomation.Concepts.Anthropic 1.1.0

AnointedAutomation.Concepts.Anthropic

Jesus is King ✝️

An optional adapter for AnointedAutomation.Concepts: it implements the core ILanguageModel port with Claude through the official Anthropic .NET SDK (13.8.0). The core package has no AI SDK and never references this one; install this package only when the language engine should talk to Claude.

Install

dotnet add package AnointedAutomation.Concepts.Anthropic

Configure

Every setting is required. There are no defaults and no fallback model: a missing value throws AnthropicConfigurationException at startup. Keep the API key in user secrets, environment variables or a secret store, never in source control. The key is never logged.

{
  "Anthropic": {
    "ApiKey": "<from your secret store>",
    "Model": "claude-opus-5-5",
    "MaxTokens": 16000,
    "TimeoutSeconds": 120,
    "MaxAttempts": 3,
    "RetryBaseDelayMilliseconds": 1000
  }
}

using AnointedAutomation.Concepts.Anthropic;

builder.Services.AddAnthropicLanguageModel(builder.Configuration.GetSection("Anthropic"));
// ILanguageModel now resolves to AnthropicLanguageModel over the named HttpClient
// AnthropicServiceCollectionExtensions.HttpClientName (add proxies or handlers there).

Build a HybridReasoner

The deterministic CanonReasoner answers first. Only when it returns Unknown does the LanguageReasoner ask Claude, once per owner, query and model; the answer is kept in the owner's private InterpretationLedger (which also caches by prompt and model, so this adapter does not cache).

using AnointedAutomation.Concepts.Reasoning;
using AnointedAutomation.Concepts.Reasoning.Hybrid;
using AnointedAutomation.Concepts.Reasoning.Language;

ILanguageModel claude = services.GetRequiredService<ILanguageModel>();
LanguageReasoner language = new LanguageReasoner(claude, new InterpretationParser(parserLogger), TimeProvider.System);
InterpretationLedger ledger = new InterpretationLedger(new InMemoryInterpretationStore(), ownerOnlyPolicy, TimeProvider.System);
HybridReasoner reasoner = new HybridReasoner(new CanonReasoner(TimeProvider.System), language, ledger, new OwnerScope("chat-123"));

Answer answer = await reasoner.Ask(query, cancellationToken);

Language answers are always Unverified and private to the owner's chat. Only the owner can promote an AI interpretation into a permanent fact: InterpretationLedger.Promote asks the injected IPromotionPolicy, and your policy must allow only the owner's Reviewer (there is no default policy).

How a call works

• One Messages API request per question. The core prompt asks for every field at once, and the reply is constrained with structured outputs (output_config.format, a JSON schema built from PromptComposer.ReplySchema). The core InterpretationParser still validates every reply and throws InterpretationRejected for anything malformed.
• 429, 408, 5xx, overload, network errors and timeouts are retried with exponential backoff (starting at RetryBaseDelayMilliseconds, capped at 30 seconds) for at most MaxAttempts calls in total, then raised as AnthropicTransientException. A Retry-After header on any retried response (429, 408, overloaded, 5xx) wins over the local backoff (still capped). The SDK's own retries are off so there is one bounded policy.
• 401 and 403 raise AnthropicConfigurationException immediately and are logged at Critical.
• Other API errors (for example 400) are logged at Error with the request id and rethrown as the SDK exception.
• Cancellation is rethrown quietly and never retried.

License

MIT. Copyright © Anointed Automation Corporation. Stewarded by Alexander Fields.