Module compaction 

Context compaction — LLM-driven conversation summarisation.

When a conversation approaches the model’s context window, older messages are summarised by a smaller/cheaper “compaction” LLM call into a structured anchored summary. This frees context tokens while preserving the semantic continuity of the conversation.

Sub-modules

• select — Turn-based message selection (head vs tail).
• prompt — Anchored summary prompt template.
• overflow — Context overflow detection.
• prune — Old tool output pruning.

Architecture

compact_if_needed() / compact_after_overflow()
  ├── overflow::is_overflow()      — detect if compaction is needed
  ├── select::select()             — split messages into head/tail
  ├── prompt::build_prompt()       — construct anchored summary prompt
  └── run_compaction()             — call compaction LLM
      └── store compaction result  — persist compaction messages

Ported from OpenCode V1/V2 compaction infrastructure.

Modules

overflow

Compact overflow detection.

prompt

Anchored summary prompt template for compaction.

prune

Tool output pruning.

select

Message selection for compaction.

Structs

CompactionCircuitBreaker

Circuit breaker that gates compaction calls after repeated failures.

CompactionResult

Result of a successful compaction.

CompactionService

Service for LLM-driven conversational context compaction.