Module context 

Context compaction for long-running conversations.

This module provides automatic context compaction to allow conversations to continue without hitting context limits. When the message history grows too large, older messages are summarized using the LLM and replaced with a compact summary.

Overview

The compaction system works as follows:

1. Monitor message history size (by token count estimation)
2. When approaching the threshold, partition messages into old and recent
3. Summarize old messages using the LLM
4. Replace history with summary + recent messages

Example

use agent_sdk::{builder, context::CompactionConfig};

let agent = builder()
    .provider(my_provider)
    .with_compaction(CompactionConfig::default())
    .build();

Configuration

Use CompactionConfig to customize compaction behavior:

• threshold_tokens: When to trigger compaction
• retain_recent: How many recent messages to keep intact
• min_messages_for_compaction: Minimum messages before considering compaction
• auto_compact: Whether to auto-compact or only on explicit request

Structs

CompactionConfig

Configuration for context compaction.

LlmContextCompactor

LLM-based context compactor.

TokenEstimator

Estimates token count for messages.

Traits

ContextCompactor

Trait for context compaction strategies.