Struct CompactionPolicy 

pub struct CompactionPolicy {
    pub strategy: Arc<dyn CompactionStrategy>,
    pub model_client: Arc<dyn ModelClient>,
    pub context_window_tokens: u64,
}

Optional compaction wiring: strategy + the model client used to run the summarize request + the resolved context-window cap. All three must travel together; without any of them the loop can’t make a useful compaction decision. AgentLoopHarness::with_compaction installs it once and the per-turn loop checks it between steps.

Fields

strategy: Arc<dyn CompactionStrategy>

model_client: Arc<dyn ModelClient>

Model client used by strategy.compact to run the summarize request. Usually the same provider as the main turn model so the Anthropic cache prefix stays hot; tests may swap in a fake.

context_window_tokens: u64

Implementations

impl CompactionPolicy

pub fn new( strategy: Arc<dyn CompactionStrategy>, model_client: Arc<dyn ModelClient>, context_window_tokens: u64, ) -> Self

Build a compaction policy from a custom strategy.

The strategy receives the full message history plus a CompactionContext containing model_client, context_window_tokens, and the current tool specs. Custom strategies may ignore the model client entirely, or use it to produce their own summaries.

pub fn summarizing( model_client: Arc<dyn ModelClient>, context_window_tokens: u64, ) -> Self

Build the default summarizing compaction policy used by the harness.

This preserves the existing compaction behavior: old oversized tool outputs are pruned first, and when pruning is insufficient the history is folded with SummarizeCompactionStrategy::default.

Trait Implementations

impl Clone for CompactionPolicy

fn clone(&self) -> CompactionPolicy

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.