Crate a2a_llm 

Provider-neutral vocabulary for chat completions, plus the providers that speak it.

LlmProvider is the port: chat_completion and chat_completion_stream over LlmRequest / LlmResponse. openai covers OpenAI and every OpenAI-compatible endpoint (OpenRouter, vLLM, llama.cpp); gemini covers Google’s API. provider_from_env picks one from the environment.

The types are deliberately not tied to A2A. ToolCall, ToolDefinition and ToolResult are the tool-calling vocabulary shared with the MCP bridge, which is why they live in their own crate rather than inside an agent framework.

decision is the second port: DecisionModel asks a decision model typed questions (yes-or-no, choice, score) and gets probabilities back instead of text.

A message’s content is MessageContent: a string, or ContentParts when it carries bytes a string cannot — an image, a PDF, a recording. Each provider renders those into the shape its API takes; see content.

Re-exports

pub use content::ContentPart;

pub use content::MessageContent;

pub use decision::DECISION_PROVIDERS;

pub use decision::Decision;

pub use decision::DecisionConfigError;

pub use decision::DecisionError;

pub use decision::DecisionModel;

pub use decision::DecisionProvider;

pub use decision::DecisionSettings;

pub use decision::Question;

pub use decision::Questions;

pub use decision::SelectedDecision;

pub use decision::answer_schema;

pub use decision::decision_from_settings;

pub use provider::LlmConfigError;

pub use provider::LlmSettings;

pub use provider::PROVIDER_ENV_VARS;

pub use provider::ReasoningPlan;

pub use provider::SUPPORTED_PROVIDERS;

pub use provider::SelectedLlm;

pub use provider::provider_from_env;

pub use provider::provider_from_settings;

pub use tool_call::PartialToolCall;

pub use tool_call::ToolCallAccumulator;

Modules

content

What a message says, when not all of it is text.

decision

Decision models: a typed answer and its probability, instead of text.

gemini

openai

provider

Centralized LlmProvider selection.

schema

What each provider takes of a tool’s JSON Schema.

tool_call

Incremental tool-call assembly for streaming LLM responses.

Structs

ChatMessage

A single message in a chat conversation.

LlmRequest

A request to an LLM provider for chat completion.

LlmResponse

A response from an LLM provider.

TokenUsage

Tokens a provider reported for one request.

ToolAnnotations

What a tool’s source says about its effects. Hints, not guarantees: MCP’s own wording, and a server can lie. None is “the source said nothing”, which is not the same as Some(false) — a tool with no read_only claim is a tool the consumer knows nothing about, not one that writes.

ToolCall

Represents a specific tool invocation requested by the LLM.

ToolDefinition

Defines a tool (function) available for the LLM to call.

ToolResult

What a tool answered: the text the model is shown, and the structured value beside it when the tool’s source returned one (MCP’s structuredContent).

Enums

FinishReason

Why the model stopped generating, normalized across providers.

LlmError

Represents an error returned by an LLM provider.

LlmStreamEvent

An event emitted during a streaming LLM response.

MessageRole

The role of the message sender.

Reasoning

What to ask a reasoning-capable model to do with its thinking.

ReasoningEffort

How hard a reasoning model should think, when reasoning is requested.

Traits

LlmProvider

Trait defining a generic LLM provider for standardizing AI integration across agents.