Module compaction 

LLM-based context compaction — summarize old messages to stay within limits.

Unlike simple sliding window (trim_messages), compaction preserves key decisions and file changes by summarizing old messages through a fast LLM.

Structs

Compactor

Compacts conversation history using LLM summarization.

Enums

CompactionError

Errors from compaction.

Functions

estimate_tokens

Estimate token count for messages (rough: chars / 4). Uses char count (not byte length) for correct non-ASCII estimation.