Module context 

Context window management — smart truncation and token counting.

The #1 engineering challenge for agents. This module provides:

• Token estimation (fast, no external deps)
• Tiered compaction (tool output truncation → turn summarization → full summary)
• Execution limits (max turns, tokens, duration)

Designed based on Claude Code’s approach: clear old tool outputs first, then summarize conversation if needed.

Structs

ContextConfig

Configuration for context management

ContextTracker

Tracks context size using real token counts from provider responses combined with estimates for messages added after the last response.

DefaultCompaction

Default 3-level compaction: truncate tool outputs → summarize turns → drop middle.

ExecutionLimits

Execution limits for the agent loop

ExecutionTracker

Tracks execution state against limits

Traits

CompactionStrategy

Strategy for compacting messages when context exceeds budget.

Functions

compact_messages

Compact messages to fit within the token budget using tiered strategy.

estimate_tokens

Rough token estimate: ~4 chars per token for English text. Good enough for context budgeting. Use tiktoken-rs for precision.

message_tokens

Estimate tokens for a single message

total_tokens

Estimate total tokens for a message list