Struct Transcript 

pub struct Transcript {
    pub meta: SessionMeta,
    pub convo: Conversation,
    pub configs: Vec<RunConfig>,
    pub config_positions: Vec<usize>,
    pub episode: Option<RunStats>,
    pub taint_timeline: TaintTimeline,
}

A transcript read once: see Session::read.

Fields

meta: SessionMeta

convo: Conversation

configs: Vec<RunConfig>

Every RunConfig recorded, in order. The first is the run the session began under; a /model switch appends another.

config_positions: Vec<usize>

How many messages preceded each entry of configs in the loaded list — parallel to it, and what Transcript::config_covering reads. A front-end writes a Config at run start, before the run’s own messages, so the config in effect at message i is the last one with a position at or below i. A summarising Rewrite clamps every earlier position to zero: the rewritten head’s original indices are claims about a list that no longer exists, and the config in flight at the rewrite — the last of the clamped ones — is the honest answer for it (messages the rewrite kept from earlier attaches were genuinely recorded under older configs, but those turns are exactly the “not comparable” case run_configs’s own doc names, and the replay fidelity caveat is the place that says so). A truncating rewrite — the failed-turn rollback, whose new list is a strict prefix of the one in hand — rewrites nothing, so its positions stay exact; see the Rewrite arm in Session::read.

episode: Option<RunStats>

Every recorded outcome, folded into the episode the session describes.

taint_timeline: TaintTimeline

The taint checkpoints, positioned against the loaded messages — the same structure Session::taint_timeline builds from a second full read, carried here because this pass already walked every record. mecha distill paid four complete read-and-parse passes per session (load, taint_timeline, then for_session’s own read and taint_timeline) for questions this one walk answers together.

Implementations

impl Transcript

pub fn config_covering(&self, message_index: usize) -> Option<&RunConfig>

The run config in effect at message message_index of the loaded list, or None for a transcript recorded before configs were kept.

This is what a replay driver should ask, not configs.first(): resuming under different flags is a normal thing to do, and replaying a later attach’s turns under the first attach’s system prompt and tool list diverges for reasons that say nothing about those turns — which a counterfactual reader then mistakes for evidence.