Struct Compaction 

pub struct Compaction { /* private fields */ }

How much of a conversation reaches the model, for every run a workspace mints.

Set with WorkspaceBuilder::with_compaction; unset, Compaction::default() applies. with_* returns a new value, so a host can keep one of these beside its other defaults and finish it differently per workspace.

Implementations

impl Compaction

pub fn with_keep_recent_tool_results(self, keep: Option<usize>) -> Self

Elides all but the keep most recent tool results from every provider request, or keeps all of them with None.

A host asks for this by number when it knows something basis does not: that its tool results are large, repetitive, and cheaply re-derived, and that the tokens are worth more than the transcript. What it costs is stated plainly — the model stops being able to see results it was shown. The request-only event reports that loss, but any turn that depends on comparing an early result against a late one still becomes unreliable.

Some(0) elides every tool result over 100 bytes, including the one that just came back.

pub fn with_auto_threshold_tokens(self, tokens: Option<usize>) -> Self

Triggers a summarizing pass once an estimated request reaches tokens, for a model whose context window basis does not know — or, with None, declines to name a number for one at all.

The estimate is mentra’s, over the serialized messages plus the system prompt, and it is not the provider’s accounting. with_auto_threshold_percent wins over this whenever the window is known, so this is the fallback a host sets for a model it cannot ask, or the only number that matters once the percentage is cleared.

None here is no longer the off switch for the whole mechanism, which is what it meant until mentra 0.24 separated the switch from the number (AutoCompactTrigger). Cleared while the percentage stands, it now asks for the posture that used to be unspellable: compact at a share of a known window, and do not compact at all for a model whose window nobody reports — rather than falling back on a figure a host had to invent precisely for the case where inventing one is worst. Clearing both numbers is the off switch, and the only spelling of it.

Off is a real posture, not a way of saying later: nothing then shortens the history ahead of time, and a conversation that outgrows the window gets exactly the recovery every other posture also gets — the provider’s own refusal triggers one compaction and one retry (see this module’s header) — rather than a second, earlier attempt basis would have made on its own guess. It suits a run that is bounded by construction — one prompt, a handful of turns — where an earlier pass would only ever cost a model call it did not need.

pub fn with_auto_threshold_percent(self, percent: Option<u8>) -> Self

Triggers a summarizing pass once an estimated request reaches percent of the model’s context window, when that window is known — or pins the trigger to with_auto_threshold_tokens’s absolute number regardless, with None.

The window is known when PreparedRun::context_window returns Some — today, a workspace whose provider is Gemini; Anthropic and the Responses transport do not report one, so this has no effect for them until they do. Values above 100 are treated as 100.

With with_auto_threshold_tokens cleared this is the whole trigger: a model that reports no window then never auto-compacts, instead of falling back on an absolute number the host declined to give. Clearing this one as well is how automatic summarizing is turned off — see there.

basis reads mentra’s own default (75) here rather than choosing one: unlike the absolute number, a percentage is not a guess about any particular model, so there is nothing basis knows that would make a different figure more honest.

pub fn with_preserve_recent_user_tokens(self, tokens: usize) -> Self

Keeps at least tokens of the most recent user turns verbatim through a summarizing pass.

What the user actually asked for is the thing a summary is least able to stand in for, which is why mentra protects a budget of it and why this is a knob rather than one of the six that stayed behind.

Trait Implementations

impl Clone for Compaction

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for Compaction

impl Debug for Compaction

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

Formats the value using the given formatter.

impl Default for Compaction

Keeps every tool result, and leaves every one of mentra’s summarizing numbers where mentra put them.

The three it leaves alone are read off mentra’s own default rather than copied into a constant here. basis has no basis for choosing any of them — the window-relative trigger is a fact about the model, and the preserved-user-text budget is a property of how mentra summarizes — so restating them as literals would mean maintaining a second opinion that can silently drift from the first.

fn default() -> Self

Returns the “default value” for a type.

impl Eq for Compaction

impl PartialEq for Compaction

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for Compaction