Enum HistoryMode 

pub enum HistoryMode {
    CacheAware,
    Rolling,
    Off,
}

How the proxy prunes old tool results from conversation history.

Provider prompt caches (Anthropic cache_control, OpenAI automatic prompt caching) bill cached prefix tokens at a fraction of the base rate but only match exact prefixes. Any mutation whose position depends on the current conversation length (a rolling window) rewrites a previously-stable message every turn, invalidating the cache from that point — turning cheap cache reads into full-price writes.

Variants

CacheAware

Prune only at frozen generation boundaries that advance in large, deterministic steps. Between jumps the request prefix is byte-stable, so provider prompt caches keep hitting. Content the client has marked with a cache_control breakpoint is never rewritten, so an advancing boundary can no longer invalidate the already-cached prefix (#448). Default.

Rolling

Legacy behaviour: summarize everything older than the last N messages. Maximum raw-token reduction, but defeats provider prompt caching.

Off

Never prune history (tool-result compression still applies — it is content-deterministic and therefore prefix-stable).

Trait Implementations

impl Clone for HistoryMode

fn clone(&self) -> HistoryMode

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for HistoryMode

impl Debug for HistoryMode

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Eq for HistoryMode

impl PartialEq for HistoryMode

fn eq(&self, other: &HistoryMode) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for HistoryMode