Binder (lncrawl/services/binder/)

A dispatch table archive_maker: Dict[OutputFormat, Callable] in service.py maps each format to a builder. EPUB, JSON, and text have native builders; every other format is convert_epub — Calibre converts from a previously built EPUB artifact (it raises ServerErrors.no_epub_file without one). available_formats derives from Calibre availability: without Calibre only the native formats exist. Zip-emitting builders are listed in requires_zip.

Builder contract (all builders): def make_x(working_dir: Path, artifact: Artifact, signal=Event(), **kwargs) -> None — write to a temp file inside working_dir, then atomically replace ctx.files.resolve(artifact.output_file). Poll signal.is_set() and raise AbortedException for cancellation. make_artifact owns the Artifact row, temp-dir lifecycle, and cleanup.

Calibre (calibre.py) has two paths: a remote convert API (posted through ctx.http.session(signal)) and the local ebook-convert executable, with configurable fallback. Options are built backend-agnostic as (flag, value) tuples and rendered per path.

Recipe — new output format:

1. Add the value to OutputFormat in lncrawl/enums.py (string enum; note name ≠ value is allowed, e.g. text → "txt"). Enum changes need a Postgres enum-sync migration — see the db-migration skill.
2. Register it in archive_maker. If Calibre supports the extension, point it at convert_epub (one line); otherwise write a builder following the contract (add to requires_zip if it emits a zip).
3. Nothing else: available_formats and the API derive from the enum + table. Per-tier format gating is separate (ctx.tier.enabled_formats).

Files: ctx.files (services/file.py) resolves artifact-relative paths under the app data dir, writes atomically, and transparently gzip-compresses text — chapter content on disk is compressed; always read it via ctx.files.load_text(...).

Translation (lncrawl/services/translators.py)

Translation runs in-process via the external lncrawl-translator package (imported as translator; sibling repo). ctx.translator (TranslationService) owns the glossary loop and persistence; the package owns engines, routing lanes, rate limits, retries, and failover.

• ctx.translator.engine — a lazily constructed translator.TranslatorService: a sync facade running the async engine router on its own event-loop thread. Its YAML config path comes from ctx.config.translator.config_file (app data dir); engines/keys/routing are edited through the mounted dashboard, not lncrawl settings. close() is guarded in ctx.destroy().
• Dashboard — the package's web UI is mounted at /api/translator by an admin-gated ASGI wrapper (server/api/translator.py, TranslatorDashboard): Bearer/?token=/cookie auth, token→cookie redirect dance, trailing-slash enforcement. It is a raw ASGI mount — app-level exception handlers and router security do NOT apply inside it.
• Calls — _translate_texts/_translate_html build request dicts and go through _invoke, which maps package errors to ServerErrors: ApiError 503 → translation_quota_exhausted, other ApiError/ValidationError → translation_failure, facade AbortedError → AbortedException. The job signal and ctx.config.translator.request_timeout are passed to every engine call.
• Glossary loop — translate_novel/volume/chapter load the stored NovelGlossary, send it with each request, and merge returned new_terms back. Chapter translation dedupes via a content hash and stores results through ctx.files.save_text.
• Detection — ctx.translator.detect_language(text) (local, no quota, no event loop) returns an ISO 639-1 code or None. fetch_novel/fetch_chapter use it to fill Novel.language when the source doesn't provide one; values are normalized via _normalize_language in services/crawler.py (drops multi/unknown, maps zh-cn→zh).

Changing engines/prompts/chunking happens in the lncrawl-translator package (sibling repo), not here — bump the dependency version in pyproject.toml to pick up a release.