Struct SummarizeCompactionStrategy 

pub struct SummarizeCompactionStrategy {
    pub trigger_fraction: f64,
    pub tail_min_messages: usize,
    pub summary_max_tokens: i32,
    pub summary_prompt: String,
    pub user_message_token_budget: u64,
}

Codex-style local compaction: send the full history to the model with a handoff-summary prompt appended, collect the reply as a checkpoint, then rebuild history as [...retained user messages, summary].

Design rationale:

• User messages are preserved verbatim (up to user_message_token_budget) because they carry precise constraints and goals that paraphrasing loses.
• Assistant / tool messages are folded into the summary — they are large but low-density and tolerate lossy compression.
• The summary is placed LAST so the model reads it as the most recent context rather than as background preamble.
• Reusing the same model client keeps the Anthropic prompt-cache prefix hot (identical system + tools on every call).

Fields

trigger_fraction: f64

tail_min_messages: usize

Minimum total message count below which compaction is skipped. Does not control retention; use user_message_token_budget for that.

summary_max_tokens: i32

summary_prompt: String

user_message_token_budget: u64

Token budget for verbatim user-message retention in the replacement history.

Implementations

impl SummarizeCompactionStrategy

pub fn with_trigger_fraction(self, fraction: f64) -> Self

pub fn with_tail_min_messages(self, n: usize) -> Self

pub fn with_summary_max_tokens(self, n: i32) -> Self

pub fn with_user_message_token_budget(self, budget: u64) -> Self

Trait Implementations

impl CompactionStrategy for SummarizeCompactionStrategy

fn should_compact( &self, messages: &[ChatMessage], context_window_tokens: u64, ) -> bool

Boolean gate. Pure (no side effects). Cheap to call every step.

fn compact<'life0, 'life1, 'async_trait>( &'life0 self, messages: Vec<ChatMessage>, ctx: &'life1 CompactionContext, ) -> Pin<Box<dyn Future<Output = Result<CompactionOutcome, CompactionError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Fold history. Caller hands over the full messages list and expects a shorter list back (typically [summary, ...tail]). Failures bubble up; agent_loop treats them as “skip this turn’s compaction” rather than failing the whole turn.

impl Default for SummarizeCompactionStrategy

fn default() -> Self

Returns the “default value” for a type.