Struct Corpus 

pub struct Corpus {
    pub rows: Vec<RunRow>,
    pub sessions_read: usize,
}

Every run the scan looked at, newest session first.

Fields

rows: Vec<RunRow>

sessions_read: usize

Sessions read, including those that contributed no rows — the denominator for “how much of the store did this answer come from”.

Implementations

impl Corpus

pub fn scan(dir: &Path, scan: &Scan) -> Result<Corpus>

Read outcomes out of the session store.

Best-effort per session, like every other reader over this store: an unreadable or torn transcript contributes nothing and does not stop the ones after it.

pub fn len(&self) -> usize

pub fn is_empty(&self) -> bool

pub fn filter(&self, keep: impl Fn(&RunRow) -> bool) -> Corpus

Keep only the rows a predicate accepts. sessions_read is preserved, because it describes the scan and not the selection.

pub fn tool_calls(&self) -> u64

pub fn tool_errors(&self) -> u64

pub fn tool_error_rate(&self) -> Option<f64>

Share of attempted calls the environment refused.

None when nothing was attempted — the denominator is zero, and a rate over no calls is undefined rather than perfect. That distinction is the one a threshold silent on zero gets wrong, which is how a trigger that stopped working entirely read as healthy.

pub fn ended_on_failed_call(&self) -> usize

Runs that decided they were done with their last call failed.

pub fn rate_of(&self, of: impl Fn(&RunRow) -> bool) -> Option<f64>

Share of runs that did. None on an empty corpus, for the reason above.

pub fn stop_causes(&self) -> BTreeMap<Option<StopCause>, usize>

How runs ended. Runs recorded before stop_cause existed, or written by a path that did not set it, count under None rather than being assumed complete.

pub fn compactions(&self) -> u64

pub fn cost_usd(&self) -> (f64, usize)

Total cost, and how many rows knew theirs. Reported as a pair because a total over partial data is a lower bound, and one that does not say so is a wrong number.

pub fn by_model(&self) -> BTreeMap<String, Corpus>

Split by model, so a rate can be read against the thing that produced it. A corpus spanning two models has no single error rate worth quoting.

Trait Implementations

impl Clone for Corpus

fn clone(&self) -> Corpus

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Corpus

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

Formats the value using the given formatter.

impl Default for Corpus

fn default() -> Corpus

Returns the “default value” for a type.