Skip to main content

CompactionStrategy

ailoop_context

Trait CompactionStrategy 

Source 
pub trait CompactionStrategy: Send + Sync {
    // Required methods
    fn name(&self) -> &'static str;
    fn compact<'life0, 'life1, 'life2, 'async_trait>(
        &'life0 self,
        messages: &'life1 [Message],
        pinned: &'life2 [bool],
        preserve_n_last: usize,
    ) -> Pin<Box<dyn Future<Output = Result<CompactionOutput, CompactionError>> + Send + 'async_trait>>
       where 'life0: 'async_trait,
             'life1: 'async_trait,
             'life2: 'async_trait,
             Self: 'async_trait;
}
Expand description

User-implementable strategy for shrinking a History’s history when History::compact_if_needed runs.

Ships with two built-in implementations: TruncateStrategy drops the oldest unpinned messages until the budget fits; SummarizeStrategy replaces the dropped prefix with a model- generated summary so context is compressed rather than lost. Use Box<dyn CompactionStrategy> to swap algorithms at runtime, or implement the trait yourself for domain-specific reductions (sliding window, importance scoring, RAG-style summarization, …).

Required Methods§

Source

fn name(&self) -> &'static str

Stable, machine-readable name of the strategy. Used by HistoryCompacted events so callers can attribute compaction to a specific algorithm in logs/metrics.

Source

fn compact<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, messages: &'life1 [Message], pinned: &'life2 [bool], preserve_n_last: usize, ) -> Pin<Box<dyn Future<Output = Result<CompactionOutput, CompactionError>> + Send + 'async_trait>>
where 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, Self: 'async_trait,

Compact messages into a smaller history.

pinned is a parallel slice of the same length as messages: pinned[i] == true marks messages[i] as “must survive”. A strategy must include every pinned message in its output (in the original relative order) and forward its true pin state in the returned CompactionOutput::pinned.

preserve_n_last is a hint: at minimum the last N messages (after walking back to a safe boundary that doesn’t strand a ToolResult from its ToolCall) should be kept verbatim.

Implementors§