conversation transcripts from the ai providers used with ghost. these are historical records with multiple speakers, including exploratory ideas and superseded proposals.

preserve original message text and order, including assistant progress updates. include the source chat title, conversation id when available, original timestamps, and the export cutoff. keep summaries and personal annotations distinct from the original messages.

include the generating model identifier/version for each assistant response when exposed by the source, along with its provider and metadata provenance. distinguish configured model labels from confirmed generating models; mark unavailable versions or model metadata explicitly rather than guessing. model changes within a chat must remain visible.

exclude internal reasoning, system and developer instructions, injected environment context, and tool payloads. retain attachment references when available; links within original messages remain as sent.

extend an existing transcript only after checking that earlier messages and annotations are preserved. saving a conversation does not make it public or turn every proposal into an endorsed decision.