Compendium.Adapters.Anthropic 1.0.0-preview.0

compendium-adapter-anthropic

Anthropic (Claude) AI provider adapter for the Compendium event-sourcing framework. Implements IAIProvider from Compendium.Abstractions.AI against Anthropic's Messages API — synchronous + streaming (SSE), with typed error mapping and opt-in prompt caching.

Built from template-compendium-adapter-dotnet per ADR-0006 (one adapter, one repo).

Install

dotnet add package Compendium.Adapters.Anthropic

// From IConfiguration (binds the "Anthropic" section)
services.AddCompendiumAnthropic(builder.Configuration);

// Or inline
services.AddCompendiumAnthropic(o =>
{
    o.ApiKey = Environment.GetEnvironmentVariable("ANTHROPIC_API_KEY")!;
    o.DefaultModel = "claude-3-7-sonnet-latest";
    o.EnablePromptCaching = true;
});

Then inject IAIProvider:

public sealed class MyService(IAIProvider ai)
{
    public Task<Result<CompletionResponse>> SummariseAsync(string text, CancellationToken ct) =>
        ai.CompleteAsync(new CompletionRequest
        {
            Model = "claude-3-5-haiku-latest",
            SystemPrompt = "You are a concise summariser.",
            Messages = new List<Message> { Message.User(text) },
        }, ct);
}

A runnable end-to-end demo (sync + streaming, prompt caching enabled) lives in samples/01-chat.

Configuration

Bind from configuration under the Anthropic section:

{
  "Anthropic": {
    "ApiKey": "sk-ant-...",
    "DefaultModel": "claude-3-7-sonnet-latest",
    "EnablePromptCaching": true
  }
}

Prompt caching (preview)

When EnablePromptCaching = true, the adapter emits a cache_control: { type: "ephemeral" } marker on the trailing system-prompt block of every request. Anthropic then caches that prefix for ~5 minutes (the ephemeral tier), so repeat calls with the same system prompt only pay the input-token cost for the delta (typically the user message + any tools).

This is most useful when:

• Your system prompt is large (≥1024 tokens for Sonnet, ≥2048 for Haiku) and stable across calls.
• You are doing high-frequency turns over the same persona/context.

Models that currently support prompt caching include the Claude 3.5/3.7/4.x Sonnet and Opus families. See Anthropic's prompt-caching docs for the authoritative list and pricing. The adapter writes the marker unconditionally when the option is on; Anthropic silently ignores it on unsupported models.

What's in scope

Out of scope (this preview)

• Provider-native tool / function calling. IAIProvider is text-only; the Compendium agent layer (Compendium.Application.StandardAgent) implements tool use through ReAct-style prompt injection, which already works against Claude. Native tool_use content blocks are deferred until the abstraction grows a tools surface.
• Vision / image content blocks. Same reason — Message.Content is string. Use a multimodal-aware surface (or a data: URI inlined in the message body) until the abstraction expands.
• Extended thinking (beta-gated reasoning blocks) — deferred behind a feature flag.
• Computer use — out of scope for v1.0.0-preview.
• AWS Bedrock / Vertex AI hosted Claude — those will live in separate adapters.

SDK choice

Hand-rolled HttpClient + System.Text.Json, mirroring compendium-adapter-openrouter. Rationale: avoid churn from official SDK pre-1.0 surfaces, reuse the Microsoft.Extensions.Http.Resilience policies already pinned by the framework, and keep the runtime graph minimal.

Repository conventions

Build & test locally

dotnet restore
dotnet build -c Release
dotnet test  -c Release --collect:"XPlat Code Coverage"

Releasing

Tag with a v prefix on main to publish to nuget.org + GitHub Packages:

git tag v1.0.0-preview.0
git push origin v1.0.0-preview.0

See docs/RELEASE.md for the full release procedure and required secrets.

License

MIT — Copyright © 2026 Sassy Solutions.