Namespace LMKit.Agents.Tools

Classes

LMFunctionAttribute

Marks an instance method as callable by LM-Kit’s tool binding and discovery.

Methods annotated with LMFunctionAttribute can be discovered and exposed as ITool via APIs such as LMFunctionToolBinder.

LMFunctionToolBinder

Discovers methods annotated with LMFunctionAttribute and binds them to ITool adapters.

This type is stateless: each call inspects the provided instance/type/assembly via reflection, validates parameter types, and returns newly constructed ITool instances.

ToolCall

Represents a single tool/function call emitted by a model.

A ToolCall is produced by parsing the model’s structured tool-call output and is then used by the runtime to dispatch the call: look up the matching ITool by Name and invoke it with ArgumentsJson.

The Id is intended for correlation across the request/response cycle: include the same identifier when returning the corresponding tool result so the model can associate each result with its originating call.

ToolCallPolicy

Per-turn policy that governs if and how the assistant may invoke registered tools.

ToolCallResult

Represents the outcome of executing a single tool call.

Instances of ToolCallResult are created by the runtime after invoking a tool and are sent back into the chat loop so the model can produce a final answer that incorporates the tool output.

The result is correlated to its originating request via ToolCallId. The payload itself is provided as JSON in ResultJson and is classified via Type.

ToolRegistry

Registry for ITool instances that can be exposed to models/agents.

Tools are stored by their stable Name (case-sensitive, using Ordinal). The registry provides read-only enumeration of registered tools and name-based lookup.

Interfaces

ITool

Contract for a model-callable tool (“function”) that an LM-Kit agent can expose to an LLM. The model discovers the tool via Name, inspects Description and InputSchema, and invokes InvokeAsync(string, CancellationToken) with JSON arguments.

Enums

ToolCallResultType

Classifies the outcome of invoking a single model-callable tool.

Consumers MUST handle unknown values defensively to remain forward-compatible with newer SDK versions (e.g., by treating unknown values as Error or logging and continuing).

ToolChoice

Controls how (or whether) the assistant may invoke registered tools during a turn.