llm-adapters — Architecture

The @mnemose/llm-adapters package provides native, multi-vendor LLM adapters behind the @freeside-collective/agent-runtime ApiClient surface. Unlike @mnemose/llm-provider (a text-only port not on the agent tool-use path), this package produces the ApiClient instances the agent loop actually streams through.

Responsibilities

Resolve an ApiClient from an LLM provider config (provider kind, secret ref, model) via createApiClient(config, deps)
Implement per-vendor adapters that translate the agent-runtime request stream into vendor SDK calls
Resolve secrets from Secret Manager via the secrets dependency and cloud credentials via cloudCredentials

Key Components

File	Role
`src/registry.ts`	`createApiClient()` — dispatches to the correct adapter by `providerKind`
`src/adapters/anthropic.ts`	Anthropic Messages API adapter
`src/adapters/vertex.ts`	Vertex AI (Gemini) adapter
`src/adapters/openai-compat.ts`	OpenAI-compatible endpoint adapter (OpenAI, OpenRouter, Azure OpenAI, …)
`src/adapters/bedrock.ts`	AWS Bedrock adapter (Anthropic on Bedrock)
`src/adapters/azure.ts`	Azure OpenAI adapter
`src/adapters/_translate.ts`	Shared request/response translation helpers
`src/config.ts`	Shared config types (`BaseLlmConfig`, `LlmProviderKind`)

Dependencies

@mnemose/domain — LlmProviderConfig Zod schemas (secret ref, model, provider kind)
@mnemose/cloud-adapters — secret resolution + cloud credentials
vendor SDKs (@anthropic-ai/sdk, OpenAI SDK, etc.) behind each adapter

Data Flow

agent loop (services/loop-host)
  ↓ createApiClient(providerConfig, { secrets, cloudCredentials })
  ↓ client.stream(request, tools) → AsyncIterable<AssistantEvent>
  ↓ vendor SDK call (Anthropic / Vertex / OpenAI / Bedrock / Azure)