scourgify

The tag-wrangler / canonizer for your fanfiction library — normalize and consolidate tags, fandoms, characters, relationships and genres in a FanFicFare-imported Calibre library. Data-driven from ~1,700 bundled generic defaults, fully customizable, audit-first and reversible.

pipx install scourgify                                # or: uv tool install scourgify  (one dependency: rich)

export CALIBRE_LIBRARY="$HOME/Calibre/fanfiction"     # folder containing metadata.db
scourgify                                             # ← the wizard: the whole lifecycle, guided

Requires Calibre installed — the tool reads via read-only sqlite and shells out to Calibre's own calibre-debug for writes. From a checkout, uv run scourgify (or uvx --from . scourgify) runs it without installing; uv handles the environment (one dependency: rich).

The wizard (no arguments) is the intended way in. It opens on a status header (book count, column health, how many books are new/changed, any pending proposal / new-tag candidates / rejects) and then asks what you want to do — run the whole guided lifecycle, or jump straight to a single task — with any unfinished work flagged inline on the menu. A single task then offers to continue to its natural next step, so you can flow onward from wherever you jumped in. On a fresh library it detects missing columns/config and runs setup first. The guided run walks the lifecycle in order — wrangle → staleness → classify → review → promote → backfill — where every stage dry-runs first, shows its report, and asks apply-all / review 1-by-1 / skip before writing. The classify stage targets only new/changed books, prices each engine for the run (public list prices), runs cloud requests 8-wide, and can compare engines on a 5-book sample before you commit; runs show a live dashboard (progress + tagged/failed/rate + throughput sparkline + rising tag candidates). Every write previews first, asks for confirmation, and auto-backs-up metadata.db.

Review 1-by-1 (--step) walks the changed books one at a time — a per-item checklist, everything pre-ticked, untick to reject — for both the wrangle stage's per-book oddities and the classify stage's AI-guessed tags. A rejected deterministic (wrangle) change is a rule bug: it's logged so scourgify overrides can fold it into a personal override that stops it recurring. A rejected classify tag is just an AI miss: dropped and logged, nothing to fix.

Each step is also a plain scriptable subcommand:

scourgify setup                                      # interactive health check + setup (FanFicFare, columns, config)
scourgify audit                                      # read-only dry-run report of every pass
scourgify apply --apply                              # write changes (Calibre CLOSED for this step)
scourgify apply --step                               # review each changed book 1-by-1 (untick items to reject)
scourgify overrides                                  # turn rejected deterministic changes into personal override rules

Everything runs under plain python3. The tool reads via read-only sqlite and, for the actual writes, shells out once to calibre-debug -e _writer.py (Calibre's API is the only fast batch-write path) — so any command that writes (apply --apply, setup creating columns, classify.py --apply) needs Calibre closed, and refuses to run while it's open. You never invoke calibre-debug yourself.

setup is the first-run wizard + re-runnable health check. It verifies, with ✓/⚠/✗ status and Y/n prompts (default-yes; --yes to auto-accept): the library; that the FanFicFare plugin is installed and configured, flagging + offering to fix the known gotchas (fandom-vs-series mapping, include_in_series:category, unprotected #genres); that every needed column exists (#fandoms, #characters, #relationships, #genres, #status, plus #updated and #wrangled for staleness / incremental classification), creating any that are missing; and writes config.toml (preserving your behavior toggles). Safe to re-run anytime.

rich is required for the wizard and powers the live dashboards/tables everywhere else; the plain subcommands still degrade to text without it (rich is try/except-imported in the core tools, so scripting/CI without rich keeps working, and _writer.py under Calibre's bundled Python needs none).

The engine reads, first to last (later wins):

• defaults/ao3/ — the bundled master taxonomy, generated from AO3's official tag dump: ~15 years of volunteer tag-wrangler knowledge as master,name,rel pair rows — universes (one name per franchise, media splits and renamed adaptations folded in: Game of Thrones (TV) → A Song of Ice and Fire), tags (93k canonical-spelling folds), characters (38k name folds), genres (3.5k synonyms). Covers every fandom above ~10 AO3 uses, not just one library. Machine-generated — never hand-edit.
• defaults/ — curated generic knowledge on top (franchise taste, junk rules, the genre allowlist, ao3_exceptions.txt for AO3 mergers deliberately not followed). Edit here to change behavior for everyone; a re-point of a generated master cascades over its whole subtree.
• config.toml — your column mapping + opinionated behavior toggles (generated by setup).
• overrides/ (optional) — your own files (same formats) that win over everything and survive pip upgrades. For personal taste that shouldn't ship.

Data from the OTW's Selective data dump for fan statisticians (2021-02-26), released for public reuse — thank you, Tag Wrangling volunteers.

⚠️ Taxonomy vintage — the AO3 layer is frozen at the 2021-02-26 dump. It's the only public dump OTW has released, so franchises that blew up after it (e.g. Baldur's Gate 3, Elden Ring, Honkai: Star Rail) aren't in defaults/ao3/ and won't auto-fold on a fresh install. The override/promote loop covers the gap — the classifier still content-tags these books, and the #fandoms/#characters values pass through untouched (they just aren't unified to a franchise canonical). When OTW publishes a newer dump, build_ao3_layer.py regenerates the whole layer.

Add a post-2021 fandom yourself — drop rows into overrides/ (they win over everything and survive upgrades; formats under Customizing):

# overrides/fandoms.csv   (alias,canonical) — unify a franchise's media splits
Baldur's Gate 3,Dungeons & Dragons
BG3,Dungeons & Dragons
# overrides/characters.csv (variant,canonical,fandom) — fold a character's abbreviations (blank = global)
Astarion Ancunín,Astarion,Dungeons & Dragons

Or just let scourgify promote adjudicate the novel tags your classify runs surface — same effect, no hand-editing.

---

FanFicFare → Calibre columns (how the linking works)

FanFicFare scrapes metadata fields from each story and writes them into Calibre columns. The mapping lives in the FFF Calibre-plugin config (stored per-library in the metadata.db preference namespaced:FanFicFarePlugin:settings → key custom_cols). scourgify's setup reads that mapping, and creates any recommended columns you're missing.

Recommended mapping (FFF metadata field → Calibre column):

FanFicFare field	Calibre column	Type	Holds
category	#fandoms	text, multiple	fandom(s) — map this to category, not series (see gotcha)
characters	#characters	text, multiple	characters
ships	#relationships	text, multiple	pairings
genre	#genres	text, multiple	genres
status	#status	text	In-Progress / Completed / …
series	Series (built-in)	series	the real site/AO3 series
numWords	#words	int	word count
numChapters	#chapters	int	chapter count
dateUpdated	#updated	datetime	last-updated date
storyUrl	#storyurl	text	source URL (also stored as the url identifier)
subject tags	tags (built-in)	—	freeform tags (what this tool normalizes most)

⚠️ The fandom-vs-series gotcha

FanFicFare's personal.ini setting include_in_series:category stuffs the fandom into FFF's series field. If your custom_cols then maps #fandoms ← series, two things break: your Series column fills with fandom names (not real series), and #fandoms is fed from that fandom-stuffed series field. Fix:

1. Remove include_in_series:category from personal.ini → series becomes the real (e.g. AO3) series.
2. Map #fandoms ← category (the true fandom field).

wrangle.py setup detects and offers to fix both, plus the protection below (the original attic/apply_fff_config.py does the same standalone). After that, real series fills in going forward and fandoms come from category.

Why fandom-as-series is especially bad: Calibre's Series is a numbered field — every book gets a series_index (A Fandom Name [1], [2], …). So fandom-as-series doesn't just duplicate the fandom, it invents a bogus ordered hierarchy: dozens of unrelated stories become "book 1, book 2… of Harry Potter," a sequence that reflects nothing real. Clearing it (see attic/apply_other.py) and mapping #fandoms ← category removes the fake ordering; real series (where the index is meaningful, e.g. a genuine 3-part AO3 series) then populate correctly.

Franchise unification (fandom granularity)

Related works in one universe (e.g. Fate/stay night, Fate/Zero, Fate/Grand Order) are distinct titles but one fandom. The bundled defaults/fandoms.csv unifies the obvious franchises to a single canonical — the Fate/Nasuverse works all map to Type-Moon (the studio/umbrella name the Nasuverse fandom uses). Prefer an English title as canonical wherever one exists (e.g. The Saga of Tanya the Evil, not Youjo Senki; Puella Magi Madoka Magica, not the romaji). This is a granularity preference: if you'd rather keep Fate/Zero separate from Fate/stay night, remove those rows from your overrides/fandoms.csv (or leave them unmapped). Note some franchises should stay split — Disney works are mostly standalone worlds (keep DuckTales, don't fold to a Disney mega-fandom), and Overlord (Game) vs Overlord (Anime) are unrelated. Curated unifications live in build_defaults.py's CURATED_FAN.

Protecting your cleanup from re-pollution

FFF's custom_cols_newonly ({column: bool}) controls overwrite-on-update: when true, FFF only writes that column if it's empty, so a metadata refresh won't clobber your normalized values. Recommended: newonly:true for #genres; leave #status writable so FanFicFare refreshes it on fetch (staleness.py re-derives the activity inference). The built-in tags column is never protected by this — so new downloads/updates re-add raw tag junk, and you re-run wrangle.py to clean it (see Maintenance).

---

Customizing

config.toml — column map + behavior toggles (all have sane, opinionated defaults):

Toggle	Default	Effect
fold_characters	true	apply abbreviation→full-name defaults
ascii_only_tags	true	transliterate non-ASCII tags to plain ASCII
au_as / crossover_as / reincarnation_as / time_travel_as	genre	put these tropes in #genres (tag to keep in tags)
fold_ratings	false	fold Erotica→Smut, Adult→Mature
keep_categories	true	keep Multi/Gen/F/M tags (false drops them)

overrides/ — drop in characters.csv, fandoms.csv, tropes.csv, junk.txt, genres_allow.txt, … (same formats as defaults/). Anything here is merged on top of the bundled defaults and wins on conflicts. This is where your preferences live — the code stays generic.

defaults/ formats:

• characters.csv — variant,canonical,fandom (blank fandom = global; set = homonym-scoped, e.g. Luke C. differs in Marvel vs PJO)
• fandoms.csv — alias,canonical
• tropes.csv — variant,canonical,route (route = tag|genre|character|fandom)
• genres_split.csv (combined,atoms), genres_canon.csv (variant,canonical), genres_allow.txt (the genre vocabulary)
• junk.txt — drop list (plain line = case-insensitive exact; re:<regex> = regex)
• ratings.txt — content-rating/warning vocabulary

---

Safety model

audit and apply compute the full new state in memory and assert no book loses its last fandom or character (backfill-before-strip), aborting without writing if that fails. A second guardrail aborts if tags would mass-shrink (>25% of assignments and >200 lost — the signature of an over-broad junk.txt rule; --force overrides after you've checked). A redundant tag is only stripped when the concept already lives in that book's structured column. audit is read-only (plain python3, fine with Calibre open); apply/setup use the Calibre API (Calibre closed). Every write automatically snapshots metadata.db to data/backups/ff_<timestamp>.db first (pruned to the last 20) and prints the path — scourgify rollback restores the newest, or scourgify rollback --list to pick one; the current db is itself snapshotted before a restore, so a rollback is reversible too (master rollback = a full "Export all Calibre data" backup).

Maintenance — after new downloads / updates

New stories arrive raw (junky subject tags, unfolded names). Order matters — deterministic cleanup first, content tagging second, because raw junk tags inflate a book's tag count and would hide it from the classifier's "sparsely tagged" targeting:

scourgify apply --apply                # 1. wrangle FIRST: junk-drop/canonicalize the new raw tags (idempotent)
scourgify staleness --apply            # 2. free: re-derive #status from #updated age (independent, any time)
scourgify classify --incremental       # 3. cheap: content-tag only new/changed books -> proposal
                                       # 4. review data/classify_proposal.csv
scourgify classify --apply             # 5. apply the reviewed tags + stamp #wrangled
scourgify promote --apply --backfill   # 6. grow the vocab from new-tag candidates, AND apply the
                                       #    promoted/aliased tags back onto the books that suggested them

Or just scourgify — the wizard runs exactly this loop, guided. Need a specific redo instead? scourgify classify --last 30 (the 30 most recently added) or --since 2026-06-01 (added or site-updated since a date). Re-running wrangle is always safe — it's idempotent and won't regress curated genres (it uses the full genres_allow.txt).

---

Content-based tagging — scourgify classify

Reads each book's description (#comments) and produces two outputs: (1) added_tags — tags chosen from the controlled vocabulary (defaults/classify_vocab.txt), which get applied; and (2) proposed_new — short novel tags not in the vocab, aggregated by frequency into classify_newtags_ranked.csv so you can review and promote the recurring ones into the vocab. Grows the tag set deliberately, without freeform noise.

scourgify classify --engine apple --limit 50        # propose -> data/classify_proposal.csv (dry-run, read-only)
scourgify classify --apply                          # add the proposed tags (Calibre CLOSED)
scourgify classify --apply --step                   # review each book's tags 1-by-1 first (untick to reject)

• --engine apple — on-device Apple Foundation Models via afm.swift (free, private; macOS 26+, Apple Intelligence). Ships as source; a swift toolchain runs it as-is, or from a checkout build the faster binary once: swiftc -O src/scourgify/afm.swift -o src/scourgify/afm. Lower quality — prone to over-tagging, so the prompt caps at --max-tags 6 and dumps (>2× cap) are rejected.
• --engine claude|openai|gemini|mistral — cloud APIs (ANTHROPIC_API_KEY / OPENAI_API_KEY / GEMINI_API_KEY / MISTRAL_API_KEY); defaults claude-haiku-4-5 / gpt-4o-mini / gemini-2.5-flash / mistral-small-latest, override with --model. Sharper; cheap.
• --bakeoff — run a few sample books through every usable engine and print the comparison, then exit (writes nothing). The quick "which engine tags my library best?" check before committing to a full run.
• Scope — which books a run touches (select.py owns this; newest-added-first, so --batch/--limit caps hit the new books first): --all = the whole library, every book regardless of tag count (a full paid pass — the wizard's classify stage offers this as an explicit "whole library" scope choice); --incremental = only new/changed books; --last N = the N most recently added; --since DATE = added or site-updated on/after DATE; no scope flag = books with < --min-tags (default 2) tags. Books whose description is too thin (<40 chars) are reported, not silently dropped — --text-fallback samples the book's own prose for them. Always dry-run until --apply.
• The allowed tag set is the hand-curated defaults/classify_vocab.txt ∪ a frequency-gated AO3 seed (defaults/classify_vocab_ao3.txt — the ~120 highest-use AO3 freeform tropes, generated by build_classify_seed.py from the OTW dump so the classifier isn't under-tagging out of the box) plus your overrides/classify_vocab.txt (a line appends a term, -term removes one — and can trim a seeded term) — editable even for a pipx/uv-tool install, where the bundled files live read-only in site-packages.
• Long runs save incrementally and resume on re-run (skip books already in the proposal; --fresh to restart). --batch N processes only N new books per run — handy for pacing API spend/rate limits. Cloud engines run --workers requests concurrently (default 8); apple is single-threaded (one on-device pipe). A spend gate asks for confirmation (or --yes) before sending more than 200 books to a cloud engine.
• --apply --step reviews the proposal 1-by-1 before writing: each book's proposed tags as a checklist you untick to reject. Accepted tags are applied and the book stamped; rejected tags are dropped and logged (an AI miss, not a rule bug); skip/quit leave a book pending for a later run.
• Proposals/outputs live in data/ (gitignored): classify_proposal.csv, classify_newtags_ranked.csv, classify_failures.csv. On --apply the proposal is archived to classify_proposal_applied_<ts>.csv, so a later apply can never re-add tags you've since hand-removed in Calibre.
• Incremental maintenance (--incremental): after new FanFicFare downloads, (re)tags only new/changed books — never classified, #updated newer than their own #wrangled marker, or re-fetched (added-date newer, which FanFicFare bumps on re-download — catching updates fetched late) — cents instead of a full pass. --apply auto-creates the marker column and stamps every processed book (tagged or not, so no-tag books aren't re-sent forever). State lives in the library (travels with metadata.db, no external file). A full cloud --fresh run is expensive; reserve it for vocab changes.

Growing the vocab — scourgify promote

After a classify run, novel tag candidates land in classify_newtags_ranked.csv. scourgify promote adjudicates each one adversarially against the master tag list: an advocate proposes promote / alias / reject, a skeptic (optionally a different engine via --verify-with openai) tries to refute a promote, and the difflib shortlist of nearest master tags grounds both. Verdicts land in data/promote_review.csv for review; scourgify promote --apply folds them into overrides/ (promotes → vocab, aliases → tropes.csv + a snap-map that stops re-proposal). Engines and keys are exactly classify's (--engine claude|openai|gemini|mistral|apple; your own API key in the env; apple is free/on-device). To grow the shipped vocab: run scourgify promote, review, then --apply (writes to overrides/classify_vocab.txt); manually copy the keeper terms into src/scourgify/defaults/classify_vocab.txt and commit that file.

Close the loop — promote --backfill. The classifier only writes vocab tags; a proposed_new candidate is never written to its book, so promoting/aliasing one grows the vocab but leaves the books that suggested it un-tagged (and they're stamped, so --incremental skips them). scourgify promote --backfill fixes that deterministically, with no API calls: it reads the book↔proposed_new record kept in the (archived) proposals and the ledger's verdicts, and applies each promoted tag (or an alias's target) onto exactly the books that first proposed it — union with their current tags, previewed and confirmed, Calibre closed. Combine with apply as promote --apply --backfill, or run it alone anytime to retro-fix past promotions.

Custom maps from your library (overrides/)

The bundled defaults/ are generic. Two helper workflows mined library-specific maps into overrides/ (gitignored): AO3-style tag clustering (overrides/tropes.csv) and fandom universe-unification (overrides/fandoms.csv, e.g. Avengers/Captain America (Movies) → Marvel, Game of Thrones (TV) → A Song of Ice and Fire). The engine loads these on top of the defaults automatically.

scourgify overrides grows these files the easy way: it reads the deterministic changes you rejected in apply --step (logged to data/rejects.csv) and synthesizes the exact identity-override lines that suppress them — a rejected fandom fold → fandoms.csv: X,X, a rejected genre canon → genres_canon.csv: X,X, and so on. Dry-run by default; --apply appends the lines (de-duped), --master targets the shipped defaults/ (maintainer, checkout only). Rejects it can't express as an additive override (junk un-drops, cross-column rescues) are listed for hand-editing rather than faked.

Repo layout

The package lives in src/scourgify/; the single scourgify command (cli.py) dispatches bare → wizard, setup/audit/apply/overrides → wrangle, classify, staleness, promote, and rollback.

• cli.py — the scourgify entry point (argv dispatcher over the tools below)
• wrangle.py — the engine: setup / audit / apply; with no command it launches the wizard
• wizard.py + ui.py — the interactive wizard and its rich terminal helpers (the one rich-required surface)
• classify.py — content-based tagging (LLM engines) · staleness.py — #status re-derivation
• common.py — shared core: library resolution, read-only sqlite + custom-column reading, config, and run_writer() (the single write funnel, with automatic backup)
• _writer.py — the only file that imports Calibre; a generic ops executor invoked under calibre-debug by run_writer(), never by hand
• defaults/ — bundled generic maps, shipped inside the package (read-only at runtime)
• Per-user files live in one central home — ~/.config/scourgify/ ($XDG_CONFIG_HOME/scourgify if set; override with $SCOURGIFY_HOME; mac + Linux): config.toml, overrides/ (your maps), and data/ (proposals/intermediates + data/backups/). An installed copy always reads/writes here, not the directory you happen to launch it from.
• build_defaults.py (repo root) — maintainer tool: regenerates defaults/ from the review maps in data/
• tests/ — uv run tests/test_core.py, no library needed
• attic/ — the original single-purpose pipeline, kept as provenance (see attic/README.md); scourgify supersedes it.

The per-library review-map CSVs in data/ are gitignored (they contain your library's actual data); only the generic defaults/ ship with the repo.

Development

develop is the integration branch (all work + PRs, rebase-merged so its history stays linear — no merge commits; CI runs the tests on Python 3.10 + 3.13). main is release-only and branch-protected — one Release vX.Y.Z merge commit per release, each arcing back to the develop commit it was cut from. Publishing is automated on release via Trusted Publishing (OIDC, no tokens): a merge to main ships to TestPyPI, a manual workflow dispatch promotes to PyPI. Tests need no Calibre or network: uv run tests/test_core.py and uv run tests/test_selection.py. Release steps live in CLAUDE.md → Branching & releases.