Glossary

Terms used throughout Axio documentation.

A

Agent

The core component that orchestrates LLM interactions. Receives messages, decides on tool calls, and returns responses.

AsyncIterator

A Python protocol for asynchronous streaming. Axio transports yield StreamEvent values via this protocol.

C

CompletionTransport

Protocol defining how Axio talks to LLM providers. Implement this to add support for new APIs.

Context

The conversation history passed to the LLM. Includes system prompt, user messages, and assistant responses.

ContextStore

Protocol for persisting conversation history. Implementations: MemoryContextStore, SQLiteContextStore.

Context Compaction

Technique for reducing context size when it approaches token limits. Uses summarization or truncation.

E

Event Stream

The flow of typed events from transport to agent. Includes tokens, tool calls, reasoning, and completion signals.

G

Guard

A permission check that runs before tool execution. Can allow, deny, or modify the handler input.

I

IterationEnd

An event signaling the end of one LLM call iteration. Contains usage statistics and stop reason.

L

LLM

Large Language Model. The AI model that powers agent reasoning (OpenAI, Anthropic, etc.).

M

MemoryContextStore

In-memory context storage. Fast but loses data on shutdown.

ModelSpec

A specification for an LLM model (name, provider, capabilities, context window).

P

PermissionGuard

Abstract class for implementing guards. Define check() method to allow/deny tool calls.

Protocol

Runtime-checkable interface (Python Protocol or ABC). Axio uses protocols for pluggability.

Parameter annotation

Type hint on a tool handler parameter. Axio reads annotations to build the JSON schema sent to the LLM. Use Annotated[T, Field(...)] from axio.field to attach descriptions, defaults, or numeric bounds.

R

ReasoningDelta

An event containing model reasoning/thinking tokens. Some providers (Anthropic) stream reasoning separately.

S

SSE

Server-Sent Events. Mechanism used by OpenAI/Anthropic APIs for streaming responses.

StreamEvent

Base type for all events in the agent loop. Includes text deltas, tool calls, reasoning, and iteration end.

Sub-agent

A child agent spawned from a parent agent. Used for parallel task execution or delegation.

T

Tool

A callable that the LLM can invoke. Combines a name, description, handler, and optional guards.

Tool handler

The executable logic for a tool. A plain async def function whose parameters define the input schema and whose body implements execution.

ToolUseStart

Event signaling the start of a tool call. Contains tool name and unique ID.

ToolInputDelta

Event containing partial JSON input for a tool call. Streamed for tools with large arguments.

Transport

The bridge between Axio and an LLM provider. Handles API calls, streaming, and authentication.

U

Usage

Token consumption statistics from an LLM call. Includes input and output token counts.

Other

GuardError

Exception raised by guards to deny tool execution. The error message is sent back to the model.

HandlerError

Exception raised by tool handlers for expected failures. Distinguishes from unexpected crashes.

to_thread()

Python asyncio function for running blocking code in a thread pool. Used for CPU-bound tools.