Autonovel export

Autonovel — Phase 4 (Export)

Typeset the completed novel and produce final deliverables.

When to use this skill

• State.json shows "phase": "export" (set by consistency-pass when cycles exhausted)
• User says "export the novel" / "build the PDF" / "make the ePub"

Prerequisites

!`cd "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" && ls chapters/*.md | wc -l && cat state.json | python -c "import json,sys; s=json.load(sys.stdin); print(s.get('phase'), s.get('chapters_drafted'))"`

Need: all chapters drafted, state.json phase is "export" (or user explicitly overriding).

For PDF: use a local xelatex/pdflatex installation or the repository's optional container toolchain. Check for a local engine:

!`which pdflatex || which xelatex || echo "no LaTeX engine found"`

If neither is found, look for compose.tools.yaml at the repository root and use the container command in Step 3. If neither native LaTeX nor Docker is available, skip PDF with a warning; ePub and transcript chunks still work.

Workflow

Step 0 — Unpaid foreshadow check

If story_state/facts.json does not exist, skip this step silently — the novel predates the story-state ledger (Module A) and has no foreshadow tracking. Otherwise run:

!`python3 ${HERMES_SKILL_DIR}/lib/story_state.py unpaid "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/story_state/facts.json"`

If any items print, list them to the user with their setup chapters before proceeding. Export continues regardless (a dangling thread may be intentional / sequel-bait) — this is a warning, not a gate.

Step 1 — Confirm title and author

Read state.json for novel_slug. Ask the user for the title (display title, may differ from slug) and author byline.

Step 2 — Content-notes front matter (if applicable)

Read profile from state.json and read the normalized selected packs from genre_context/resolved.json's selected_pack_ids. Let Intimacy be selected only when "intimacy" is in that list; do not re-resolve pack selection from display tags. A non-empty profile.content_tags still implies Intimacy because the resolver records "intimacy" in selected_pack_ids for that profile. If profile.rating is M or E, or Intimacy is selected, build a content-notes line:

Rating: <rating>. Tags: <include-tags>. Content: <content_tags>.

<include-tags> remains the original profile.tags display list with only the -tag exclusion entries dropped. Do not replace it with normalized tags, pack IDs, or only pack-resolving tags; unknown positive display tags remain visible. This is exactly the export-labeling rule from ${HERMES_SKILL_DIR}/../autonovel/genres/special/intimacy/guardrails.md. <content_tags> is profile.content_tags joined verbatim; include the Content: <content_tags>. part ONLY when profile.content_tags is non-empty, and drop that part entirely (line ends at Tags: <include-tags>.) when it is empty. The line lands in each format through a concrete hook:

• PDF — the tex template has a <CONTENT_NOTES> placeholder on the copyright/colophon page. Fill it with the line in Step 3, exactly as you fill <NOVEL_TITLE> and <NOVEL_AUTHOR>.
• ePub — pass the Step 2 line to build_epub.py via its optional --content-notes "Rating: <rating>. Tags: <include-tags>. Content: <content_tags>." flag in Step 4 (drop the Content: part when content_tags is empty). It becomes a front-matter page before chapter 1 plus dc:description metadata. Omit the flag when notes don't apply.
• Audiobook — write the line to $AUTONOVEL_WORKSPACE/$NOVEL_SLUG/export/audiobook/000_content_notes.txt (the 000 prefix sorts it first, so any TTS pipeline reads it as a preamble). Skip the file when notes don't apply.

When the rating is G or T and no intimacy pack is loaded, skip the ePub flag and the audiobook file — but still fill the tex <CONTENT_NOTES> placeholder in Step 3, with an empty string (an empty placeholder renders nothing; a literal <CONTENT_NOTES> left unfilled would leak into the PDF).

Step 3 — Build the PDF (if LaTeX available)

Copy the template:

!`cp ${HERMES_SKILL_DIR}/templates/novel.tex.template "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/export/novel.tex"`

Fill <NOVEL_TITLE>, <NOVEL_AUTHOR>, and <CONTENT_NOTES> placeholders (<CONTENT_NOTES> gets the Step 2 line, or an empty string when content notes don't apply).

Run the builder:

!`python3 ${HERMES_SKILL_DIR}/lib/build_tex.py --workspace "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" --out "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/export/novel.pdf"`

If it errors, read references/typesetting-checklist.md for common LaTeX issues (missing packages, Unicode in chapter text, etc.).

For the repository's containerized XeLaTeX path, set AUTONOVEL_PROJECT to the absolute per-novel workspace path, then run from the repository root:

docker compose -f compose.tools.yaml build latex
docker compose -f compose.tools.yaml run --rm latex --workspace /novel --out /novel/export/novel.pdf

The project mount is writable so the PDF and intermediate TeX files land in the novel's own export/ directory. The repository mount is read-only.

Step 4 — Build the ePub

!`python3 ${HERMES_SKILL_DIR}/lib/build_epub.py --workspace "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG" --out "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/export/novel.epub" --title "<TITLE>" --author "<AUTHOR>"`

If Step 2 produced a content-notes line, append --content-notes "Rating: <rating>. Tags: <include-tags>. Content: <content_tags>." to the command above (drop the Content: part when profile.content_tags is empty — it matches the Step 2 line exactly).

Step 5 — Audiobook chunks

For each chapter:

!`python3 ${HERMES_SKILL_DIR}/lib/chunk_audio.py --chapter "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/chapters/ch_NN.md" --out-dir "$AUTONOVEL_WORKSPACE/$NOVEL_SLUG/export/audiobook/ch_NN"`

Chunks land as .txt files. If the user has a local TTS pipeline:

• They can shell out: for chunk in chapters/ch_NN_chunk_*.txt; do my_tts < $chunk > ${chunk%.txt}.mp3; done
• The skill doesn't run TTS itself; chunking is the deliverable.

The optional f5-tts Compose profile uses the official ghcr.io/swivid/f5-tts:main image. This integration is experimental and is not part of the export acceptance gate. It requires an NVIDIA-compatible Docker runtime plus a user-supplied reference recording and exact transcript; AutoNovel bundles no voice. Set AUTONOVEL_PROJECT and F5_TTS_REFERENCE_DIR, then invoke one chunk explicitly:

docker compose -f compose.tools.yaml --profile tts run --rm f5-tts \
  --model F5TTS_v1_Base \
  --ref_audio /reference/voice.wav \
  --ref_text "Exact transcript of the reference recording." \
  --gen_file /novel/export/audiobook/ch_01/ch_01_chunk_001.txt \
  --output_dir /novel/export/audiobook/rendered

Review generated audio before processing a full manuscript.

Step 6 — Final state update

Set state.json phase to "done". Commit. Report final stats: word count, chapter count, score history, file paths to PDF/ePub/audio chunks.

Handoff

Pipeline complete. Suggest autonovel-prose-review if the user wants further polish on any specific chapters.

See also

• references/typesetting-checklist.md — common LaTeX issues and fixes
• templates/novel.tex.template — the LaTeX template (edit to customize layout)