llm-provider — Architecture

The @mnemose/llm-provider package defines the abstract LLMProvider port and built-in text-only implementations (Anthropic, OpenAI-compatible, OpenCode). It is the simpler LLM surface — it is not on the agent tool-use path (that is @mnemose/llm-adapters); it powers standalone text completion where tool streaming is not needed.

Responsibilities

Define the LLMProvider port (complete(request) → AsyncIterable<chunk>)
Provide createLLMProvider(config) factory and resolveProviderFromEnv() resolver
Implement Anthropic, OpenAI-compatible, and OpenCode text completion providers

Key Components

File	Role
`src/ports/provider.ts`	`LLMProvider` interface + `LLMProviderConfig` / `LLMProviderMetadata` types
`src/factory.ts`	`createLLMProvider(config)` — dispatches to the correct provider
`src/resolver.ts`	`resolveProviderFromEnv()` / `resolveProviderMetadataFromEnv()` — env-driven selection
`src/providers/anthropic.ts`	Anthropic text completion adapter
`src/providers/openai-compat.ts`	OpenAI-compatible text completion adapter
`src/providers/opencode.ts`	OpenCode text completion adapter

Dependencies

zod
vendor SDKs behind each provider implementation

Data Flow

caller
  ↓ resolveProviderFromEnv() → LLMProvider (or null)
  ↓ provider.complete({ messages }) → AsyncIterable<{ content }>
  ↓ vendor SDK text completion call