Export retention

Every book export writes the newest artifact bundle to its normal per-book folder. Before the write, the existing bundle is copied into that book's Archives\v<N>\ directory, where <N> is the version parsed from the artifact filename (or the node's current version for metadata without a version in its name). This applies to DOCX, EPUB, PDF, TXT, Markdown, HTML, JSON, descriptions, synopses, keywords, and cover images.

The live folder is staging for the newest export. Previous versions are never pruned by an export; the archive is permanent and belongs to that book. The current cover.jpg remains in the live folder after archiving so exports with automatic cover generation disabled do not lose author art.

Where rendering lives

Rendering is implemented once, in the shared MindAttic.Export library (github.com/mindattic/MindAttic.Export), consumed as a package from lib/local-packages. Prose keeps everything that needs the database — the read gate, the beat walk and chapter spine, the glossary, ExportPathResolver, the version bump, the press record and the ArchivedBooks snapshot — and hands a renderer-neutral manuscript to the library's docx, epub, pdf, txt and Markdown renderers. The archive pass above (ExportCleanupService), path sanitising and ProseInline are the library's too.

• prose --export-node --slug X --preview <dir> renders the book's current version into a scratch folder, read-only and reproducibly (fixed EPUB id/timestamp and PDF metadata). The read gate still applies: an unread book previews its Markdown only. Use it to verify any renderer change against the whole corpus before pressing; the 2026-10-08 migration was verified this way on all 36 GLMZ books.
• prose --export-book-report --slug X [--formats md,txt,docx,pdf] writes {CODE}_BookReport.md as before and renders the same report as .txt, .docx and .pdf beside it (Letter document style). Report files are overwritten in place, as the .md always was.

Where writing lives

Every user-facing file Prose exports — reports, data dumps (--sql-export, canon json/zip, KDP transfer, universe interchange, barks, personas), the legacy BookExportService formats, the static HTML site, episode scripts, audiobook and episode audio, BookTok clips, DALL·E images, the KDP staging copy — is written through the library's ArtifactWriter, via ProseArtifacts (src/Prose.Core/Services/ProseArtifacts.cs). Its Overwrite preset reproduces what these sites always did: replace the file in place under its exact name, UTF-8 without a BOM (OverwriteWithBom for the reader-QA report, which has always carried one). The write is atomic: a temp file beside the target is moved into place, so a failed or cancelled export leaves the previous file, never a partial one. Renderers that need a path (QuestPDF) render to that temp path; ffmpeg, which picks its muxer from the output extension, still renders into its scratch folder and the finished file is copied into place. Version archiving stays where it was (ExportCleanupService before a bundle is written).