interlens.message¶

A single committed turn in a conversation.

The transcript is stored canonically and author-centric: a message records who said what, and is deliberately agnostic to the assistant/user role distinction. That role mapping is a per-participant view concern (see Transcript.render_roles), not a property of the message itself — so one transcript can be rendered from every participant's perspective without duplicating state.

author is the participant's name (a string), never the Participant object. This keeps transcripts trivially JSON-serializable and valid even when no models are loaded (e.g. when scoring saved transcripts).

content is the committed, visible text — the only field that is authoritative for rendering history back into a model. Anything else a generation produced (parsed <think> reasoning, the raw completion, tool call/result trails, per-token logprobs) lives in metadata under neutral keys, so hidden generated text is never silently promoted into what other participants see.