🔌 The LLM ABC

All provider traffic flows through a framework-owned LLM abstract base class. The Runner never imports a provider SDK directly. Provider code lives in src/augments/adk/llms/<provider>/ and is the only place allowed to import litellm, anthropic, openai, google.genai, etc.

Shape

class LLM(abc.ABC):
    @abc.abstractmethod
    async def acomplete(
        self,
        messages: Sequence[LLMInputContentItem | RunItem],
        *,
        tools: Sequence[Tool] | None = None,
        config: LLMConfig | None = None,
        stream: bool = False,
    ) -> LLMResponse | AsyncIterator[LLMStreamEvent]: ...

• Input is Layer 1 (or Layer 3 history materialised as Layer 1) — framework-owned.
• Output is LLMResponse (parts-based union of TextPart, ThinkingPart, ToolCallPart) when stream=False, or an AsyncIterator[LLMStreamEvent] when stream=True.
• Config is provider-agnostic. Provider-specific configs subclass it (see below).

Why a framework-owned ABC

Adopting OpenAI's Model / AnyLLMModel as the canonical type would cost three conversion hops per turn (developer → OpenAI shape → litellm shape → provider shape) and would leak openai.types.* into every non-OpenAI provider implementation. The framework-owned ABC means one conversion per direction, inside each provider module.

Available implementations

Each provider implementation owns its own wire layer (ChatCompletion*, Anthropic*, etc.) and converts in/out on the boundary.

Config layering

classDiagram
  class LLMConfig
  class LiteLLMConfig
  class AnthropicConfig
  class OpenAIResponsesConfig
  class OpenAIChatCompletionsConfig
  class GeminiConfig
  LLMConfig <|-- LiteLLMConfig
  LLMConfig <|-- AnthropicConfig
  LLMConfig <|-- OpenAIResponsesConfig
  LLMConfig <|-- OpenAIChatCompletionsConfig
  LLMConfig <|-- GeminiConfig

LLMConfig declares only provider-agnostic fields (model, temperature, max_tokens, retry / cache strategy hooks). Provider subclasses add typed fields for that provider's surface (Anthropic hosted-tool params, OpenAI Responses API knobs, Gemini-specific options, etc.).

Invariants

• The Runner MUST NOT import a provider SDK.
• A provider implementation MUST NOT depend on another provider's module.
• LLMConfig MUST NOT reference any provider name.
• Wire-type TypedDicts MUST NOT be converted to @dataclass / BaseModel.