Message containing tool execution result. This is used to send tool results back to the model after it has made a tool call. - OpenAI: Serialized a

Dependency injection marker, similar to FastAPI's Depends. The dependency callable will be invoked when the tool is called, and its return value

Exception raised when a model provider returns an error.

Definition of a tool that can be called by the model. This is a provider-agnostic representation that gets serialized appropriately for each

A wrapper around Google's Gemini chat model using the genai client. This class provides tool calling support for Google Gemini models.

Response from a chat model invocation. For tool calling workflows: - If the model wants to call tools, `tool_calls` will be populated an

Usage information for a chat model invocation.

A wrapper around AsyncOpenAI that implements the BaseChatModel protocol. This class provides tool calling support for OpenAI models. Ex

Result of a compaction operation. Attributes: compacted: Whether compaction was performed. original_tokens: Token count b

Emitted when the agent produces its final response.

Emitted when the assistant produces text content.

Simple agentic loop that manages tool calling and message history. The agent will: 1. Send the task to the LLM with available tools

Pricing information for a model

Exception raised when a model provider returns a rate limit error.

Emitted when a tool returns a result.

Summary of token usage and costs

Cached pricing data with timestamp

Configuration for the compaction service. The compaction service monitors token usage and automatically summarizes conversation history when

Service for managing conversation context through compaction. The service monitors token usage after each LLM response and triggers compactio

Simulated database with some user data.

Emitted when the agent injects a hidden user message (ex: incomplete todos prompt). Hidden messages are saved to history and sent to the LLM but not

Usage statistics for a single model

Usage tokens for a single model

Raised when a path escapes the sandbox.

Emitted when a step completes.

Emitted when the agent starts a logical step (tool execution group).

Test serializer handles tool calls correctly.

Emitted when the model produces thinking/reasoning content.

Service for tracking token usage and calculating costs

Token cost

Token usage tracking for compaction decisions. Attributes: input_tokens: Number of input tokens. output_tokens: Number of

Single token usage entry

Wrapper class for a tool function with its metadata and definition. This class is created by the @tool decorator and provides: - Automat

Emitted when the assistant calls a tool.

Serializer for converting between custom message types and Anthropic message param types.

Protocol defining the interface for chat models. All LLM implementations (OpenAI, Anthropic, Google, etc.) must implement this protocol.

A wrapper around Anthropic's chat model.

A class for to interact with any provider using the OpenAI API schema. Args: model (str): The name of the OpenAI model to use.

Document content block for PDFs.

Redacted thinking content block. Used when thinking content has been redacted by the model (Anthropic). Must be preserved in conversation history.

Thinking content block for extended thinking / reasoning models. Used by: - Anthropic Claude models with extended thinking enabled - Can be includ

Developer-provided instructions for reasoning models. With OpenAI o1 models and newer, `developer` messages replace the previous `system` messages.

Source for a document attachment (base64 encoded).

Serializer for converting messages to Google Gemini format.

Emitted when a message is complete.

Emitted when a new message starts (user or assistant).

Serializer for converting between custom message types and OpenAI message param types.

Sandboxed filesystem context. All file operations are restricted to root_dir.

Exception raised when a task is completed via the done tool. This provides explicit task completion signaling instead of relying on the absen

Integration tests that actually call Gemini API.

Base class for all message types