Truncate text to max_length characters, adding ellipsis if truncated.

Invoke the model with the given messages and optional tools. Args: messages: List of chat messages tools: Op

Convert a list of BaseMessages to Google format, extracting system message. Google handles system instructions separately from the c

Serialize cache control.

Check if we're in debug mode based on environment variables.

Resolve and validate a path is within the sandbox.

Check token usage and compact if threshold exceeded. The threshold is calculated dynamically based on the model's context window. Ar

Create a no-op decorator that accepts all lmnr observe parameters but does nothing.

Create a new sandbox context, defaulting to a temp directory.

Resolve the dependency, handling both sync and async callables. Args: overrides: Optional dict mapping original dependency functions to

Fetch pricing data from LiteLLM GitHub and cache it with timestamp

Transform schema for OpenAI strict mode: all properties must be required.

Convert a text content part to Anthropic's TextBlockParam.

Invoke the model with the given messages and optional tools. Args: messages: List of chat messages forming the conversation.

Create the most optimized schema by flattening all $ref/$defs while preserving FULL descriptions and ALL action definitions. Also ens

Create a decorator for async generators that wraps them in a Laminar span.

Destroy old ephemeral message content, keeping the last N per tool. Tools can specify how many outputs to keep via _ephemeral attribute:

Execute a single tool call and return the result as a ToolMessage.

Format image URL for display, truncating if necessary.

Generate a summary of what was accomplished when max iterations is reached. Uses the LLM to summarize the conversation history and actions ta

Hook for subclasses to check for incomplete todos before finishing. This method is called when the LLM is about to stop (no more tool calls i

Invoke the LLM with current messages and tools. Includes retry logic with exponential backoff for LLM errors

Convert an image content part to Anthropic's ImageBlockParam.

Serialize content for system messages (text only).

Convert ToolDefinitions to Google's tool format.

Returns a genai.Client instance. Returns: genai.Client: An instance of the Google genai client.

Get pricing information for a specific model

Get the compaction threshold for a specific model.

Get summary of token usage and costs (costs calculated on-the-fly)

Initialize the service by loading pricing data

Send a message to the agent and stream events as they occur. Yields events for each step of the agent's execution, providing

Analyze function signature to extract parameters and dependencies.

Build a stable cache key for system/tool definitions.

Clean cache settings so only the last cache=True message remains cached. Because of how Claude caching works, only the last cache message mat

Create a cached content resource for stable system/tools.

Extract screenshot base64 from a tool message if present. Browser tools may return ContentPartImageParam with screenshots. This metho

Extract summary content from <summary></summary> tags. If tags are not found, returns the original text. Args: text:

Extract text content from Anthropic response.

Extract text content from Google response.

Extract thinking and redacted thinking from Anthropic response. Returns: Tuple of (thinking_content, redacted_thinking_data)

Extract tool calls from Anthropic response.

Extract tool calls from Google response.

Extract tool calls from OpenAI response.

Try to find model in pricing data using various name formats.

Find the most recent valid cache file

Convert a Pydantic model to a Gemini-compatible schema. This function removes unsupported properties like 'additionalProperties' and

Prepare client parameters dictionary.

Prepare client parameters dictionary.

Prepare client parameters dictionary.

Prepare client parameters dictionary for invoke.

Extract parameter description from docstring if available.

Extract stop_reason from Google response.

Convert our tool_choice to Anthropic's format.

Convert our tool_choice to Google's format.

Convert our tool_choice to OpenAI's format.

Check if the URL is a base64 encoded image.

Check if a specific cache file is valid and not expired

Load pricing data from a specific cache file

Load pricing data from cache or fetch from GitHub

Ensure all properties are required for OpenAI strict mode

Transform a single property for strict mode, recursively handling nested objects.

Parse a base64 data URL to extract media type and data.

Prepare messages for summarization. This removes tool_calls from the last assistant message to avoid API errors (tool_use requires to

Convert a Python type to JSON schema type.

Select prompt cache retention based on model support.

Serialize content for assistant messages (text and refusal allowed).

Serialize content to Anthropic format.

Convert a document content part to Anthropic's DocumentBlockParam.

Serialize content to a string.

Serialize the tool result to string or content parts. Returns content parts (list) directly for multimodal results, otherwise seriali

Convert tool calls to Anthropic's ToolUseBlockParam format.

Convert a ToolMessage to Anthropic's ToolResultBlockParam.

Convert a ToolMessage to Google's FunctionResponse Part.

Serialize content for tool messages (text only, images converted to text description).

Serialize tool result content for Anthropic.

Convert ToolDefinitions to Anthropic's tool format. Only the last N tools get cache_control to stay within Anthropic's limit of 4 cache block

Convert ToolDefinitions to OpenAI's tool format.

Serialize content for user messages (text and images allowed). No native PDF/document support, so docs converted to a text placeholder.

Add token usage entry to history (without calculating cost)

Check token usage and compact if threshold exceeded. This is the main entry point for the compaction service. It checks if compaction

Clear the message history and token usage.

Perform compaction on the message history. This method: 1. Prepares the messages for summarization (removing pending tool calls)

Execute the tool function with the given arguments. Handles: - Dependency injection resolution (with optional overrides)

Create TokenUsage from ChatInvokeUsage.

Returns an AsyncAnthropic client. Returns: AsyncAnthropic: An instance of the AsyncAnthropic client.

Returns an AsyncOpenAI client. Returns: AsyncOpenAI: An instance of the AsyncOpenAI client.

Factory function to create LLM instances from string names with API keys from environment. Args: model_name: String name like 'azure

Get the context window limit for a model.

Get usage summary for the agent. Returns: UsageSummary with token counts and costs.