DazedTL/modules/wolfdawn.py
2026-07-09 13:35:12 -05:00

478 lines
19 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

"""Translate WolfDawn extraction JSON.
This module consumes the JSON produced by the vendored WolfDawn ``wolf`` CLI
(``strings-extract`` / ``names-extract``) and fills in the ``text`` fields with
translations, leaving ``source`` untouched so WolfDawn's inject drift-guard can
match each line back to its original.
Supported document ``kind``s (all share the ``{source, text}`` leaf pattern):
map / common -> scenes[].lines[]
db -> groups[].lines[]
gamedat -> lines[]
txt -> lines[]
txt-dir -> files[].lines[]
names -> names[]
Only entries whose ``source`` contains target-language (Japanese by default)
text are sent to the model; everything else keeps ``text == source`` so inject
is a no-op for it. When ``IGNORETLTEXT`` is on (default), entries whose ``text``
is already translated are skipped so Phase 2 can resume a partial run without
retranslating completed lines. Skip looks past Japanese nameplates on line 1
and Japanese embedded only in WOLF control codes (e.g. ``\\r[...]`` ruby).
names.json safety: WolfDawn tags every value name with a static ``safety`` badge
(``safe``, ``refs``, or ``verify``). For ``kind == "names"`` documents, only
entries whose badge is ``safe`` are translated; ``refs`` and ``verify`` names and
legacy entries without a badge keep ``text == source``. After translating
names.json, short name-like entries are harvested into ``vocab.txt`` (grouped by
``note``, with bilingual headers). After foundation DB sheets are translated,
short label fields (map names, titles, etc.) are merged into the same glossary
so later phases keep place / term names consistent without overwriting names.json.
Database sheet filter (Step 4): optional ``wolfDbIncludeGroups`` or
``wolfDbIncludeTiers`` in ``.env`` limit which ``kind: db`` lines are collected.
When unset, all database lines are translated.
Speakers: WolfDawn tags each line with ``speaker`` / ``speaker_src``. For the
first-line formats (``literal_line1`` / ``literal_line1_lowconf``) the speaker
name is baked into line 1 of ``source``. Those lines are reshaped into the shared
``[Speaker]: line`` convention (which the prompt already translates) and restored
to WOLF's native ``Speaker\nline`` layout on write-back. See ``util.speakers``.
Detection is WolfDawn's, so the reliable nameplate (``literal_line1``) is always
reshaped; only the low-confidence guess (``literal_line1_lowconf``) is gated by a
per-game, AI-recommended setting from the workflow.
Layout restore: after each file is written to ``translated/``, ``wolf
layout-restore`` copies unambiguous positional whitespace pads from ``source``
onto ``text`` (status-card ``<pad>\\nNAME`` fields the model often strips).
"""
import json
import os
import re
import threading
import time
import traceback
from colorama import Fore
from tqdm import tqdm
from util.paths import PROMPT_PATH, VOCAB_PATH
from util.translation import (
TranslationConfig,
translateAI as sharedtranslateAI,
getPricingConfig,
calculateCost,
)
from util import speakers as wolf_speakers
from util import vocab as wolf_vocab
from util import wolfdawn
from util.wolfdawn import codes as wolf_codes
from util.wolfdawn import db_classify as wolf_db
from util.wolfdawn import names as wolf_names
# Globals (mirror the other engine modules; populated from .env at import time)
MODEL = os.getenv("model")
TIMEOUT = int(os.getenv("timeout"))
LANGUAGE = os.getenv("language").capitalize()
PROMPT = PROMPT_PATH.read_text(encoding="utf-8")
VOCAB = VOCAB_PATH.read_text(encoding="utf-8")
LOCK = threading.Lock()
MAXHISTORY = 10
ESTIMATE = ""
TOKENS = [0, 0]
MISMATCH = [] # Files that hit a length-mismatch during translation
FILENAME = None
# Regex - default matches Japanese (kanji, kana, full-width forms).
LANGREGEX = r"[一-龠ぁ-ゔァ-ヴーa---\uFF61-\uFF9F]+"
# Skip lines whose ``text`` is already translated. WolfDawn keeps Japanese in
# ``source`` forever, so skip-translated must look at ``text``, not ``source``.
# Set False for a forced full retranslate.
IGNORETLTEXT = True
# Control codes that can embed Japanese inside an otherwise-translated line
# (ruby ``\r[kanji,kana]``, colour / font indexes, ``\cself[n]``, etc.).
_WOLF_CODE_RE = re.compile(
r"\\(?:r\[[^\]]*\]|c(?:self)?\[[^\]]*\]|[A-Za-z]+\[[^\]]*\]|[A-Za-z])"
)
# Speaker handling: for first-line-speaker formats, reshape the line into the
# shared "[Speaker]: line" transport before translating and restore WOLF's
# native "Speaker\nline" layout on write-back. Which formats are reshaped is
# configurable from the workflow (data/wolf_speakers.json).
SPEAKER_CONFIG = wolf_speakers.load_config()
# Pricing / batching from the configured model
PRICING_CONFIG = getPricingConfig(MODEL)
INPUTAPICOST = PRICING_CONFIG["inputAPICost"]
OUTPUTAPICOST = PRICING_CONFIG["outputAPICost"]
BATCHSIZE = PRICING_CONFIG["batchSize"]
FREQUENCY_PENALTY = PRICING_CONFIG["frequencyPenalty"]
# tqdm / progress globals (PBAR is polled by util/subprocess_runner.py)
BAR_FORMAT = "{l_bar}{bar:10}{r_bar}{bar:-10b}"
POSITION = 0
LEAVE = False
PBAR = None
TRANSLATION_CONFIG = TranslationConfig(
model=MODEL,
language=LANGUAGE,
prompt=PROMPT,
vocab=VOCAB,
langRegex=LANGREGEX,
batchSize=BATCHSIZE,
maxHistory=MAXHISTORY,
estimateMode=False,
)
def _batch_phase() -> str:
"""Current Anthropic batch phase from the GUI subprocess env (``collect`` / ``consume``)."""
return (os.getenv("BATCH_PHASE") or "").strip().lower()
def handleWolfDawn(filename, estimate):
"""Entry point used by the CLI/GUI dispatchers. Returns a summary string or 'Fail'."""
global ESTIMATE, TOKENS, FILENAME, SPEAKER_CONFIG, VOCAB
ESTIMATE = estimate
FILENAME = filename
# Re-read workflow-configured settings so edits made this session take effect
# even when translation runs in-process (the module import is cached).
SPEAKER_CONFIG = wolf_speakers.load_config()
# Reload the glossary so a later phase (DB text / dialogue) picks up names
# that an earlier Phase 0 (names) harvested into vocab.txt.
VOCAB = VOCAB_PATH.read_text(encoding="utf-8")
TRANSLATION_CONFIG.vocab = VOCAB
start = time.time()
translatedData = openFiles(filename)
# Batch collect only queues API requests; text is still Japanese. Writing
# translated/ here would overwrite prior work with source echoed back.
# Real English is written on the consume pass (or a live non-batch run).
if not estimate and _batch_phase() != "collect":
out_path = "translated/" + filename
try:
with open(out_path, "w", encoding="utf-8", newline="\n") as outFile:
json.dump(translatedData[0], outFile, ensure_ascii=False, indent=4)
except Exception:
traceback.print_exc()
return "Fail"
# Restore positional whitespace pads the model dropped (status-card
# ``<pad>\nNAME`` fields, etc.). WolfDawn edits the JSON in place.
try:
res = wolfdawn.layout_restore(out_path)
if not res.ok:
tqdm.write(
Fore.YELLOW
+ f"{filename}: layout-restore exited {res.returncode}"
+ (f" ({res.stderr.strip()})" if res.stderr.strip() else "")
+ Fore.RESET
)
else:
fixed = wolfdawn.parse_layout_restore_counts(res.stdout, res.stderr)
if fixed:
tqdm.write(
Fore.CYAN
+ f"{filename}: layout-restore fixed {fixed} line(s)"
+ Fore.RESET
)
except Exception:
traceback.print_exc()
end = time.time()
tqdm.write(getResultString(translatedData, end - start, filename))
with LOCK:
TOKENS[0] += translatedData[1][0]
TOKENS[1] += translatedData[1][1]
totalString = getResultString(["", TOKENS, None], end - start, "TOTAL")
if len(MISMATCH) > 0:
return totalString + Fore.RED + f"\nMismatch Errors: {MISMATCH}" + Fore.RESET
return totalString
def openFiles(filename):
"""Load the extraction JSON, translate it in place, and return [data, tokens, error]."""
with open("files/" + filename, "r", encoding="utf-8-sig") as f:
data = json.load(f)
kind = data.get("kind")
if kind not in ("map", "common", "db", "gamedat", "txt", "txt-dir", "names"):
raise NameError(
f"{filename}: unrecognised WolfDawn document (kind={kind!r}). "
"Expected a strings-extract or names-extract JSON."
)
return parseDocument(data, filename)
def collectEntries(data):
"""Return the list of leaf {source, text} dicts for a WolfDawn document.
The returned dicts are live references into ``data`` so mutating ``text``
updates the document that gets written back out.
"""
kind = data.get("kind")
entries = []
if kind in ("map", "common"):
for scene in data.get("scenes") or []:
for line in scene.get("lines") or []:
entries.append(line)
elif kind == "db":
include_tiers, include_groups = wolf_db.load_db_filter_config()
json_file = wolf_db.json_file_from_doc(data)
for group in data.get("groups") or []:
type_name = str(group.get("typeName") or "")
if not wolf_db.group_matches_filter(
json_file, type_name, include_tiers, include_groups
):
continue
for line in group.get("lines") or []:
entries.append(line)
elif kind in ("gamedat", "txt"):
for line in data.get("lines") or []:
entries.append(line)
elif kind == "txt-dir":
for fileDoc in data.get("files") or []:
for line in fileDoc.get("lines") or []:
entries.append(line)
elif kind == "names":
for name in data.get("names") or []:
entries.append(name)
return entries
def _text_check_body(text: str, speaker_src: str = "") -> str:
"""Body text used for skip-translated checks (nameplates / codes ignored).
Already-translated WOLF dialogue often keeps a Japanese speaker name on
line 1 (``司祭\\nSorry to keep you...``) and may embed Japanese only inside
control codes like ``\\r[我,わ]``. Those lines are finished translations.
"""
if not isinstance(text, str):
return ""
_prefix, rest = wolf_speakers.split_window_prefix(text)
if "\n" in rest:
first, body = rest.split("\n", 1)
# Known first-line speaker formats, or a short nameplate-like first line.
if (
speaker_src in wolf_speakers.FIRSTLINE_SRCS
or speaker_src in ("ui", "narration")
or (0 < len(first.strip()) <= 20 and re.search(LANGREGEX, first))
):
rest = body
return _WOLF_CODE_RE.sub("", rest)
def _text_still_needs_translation(entry) -> bool:
"""True when ``text`` still needs the model (empty, Japanese, or identical)."""
txt = entry.get("text")
if not isinstance(txt, str) or not txt.strip():
return True
src = entry.get("source")
# Fresh / unfinished extract: text still mirrors Japanese source.
if isinstance(src, str) and txt == src:
return True
# Fully English (no Japanese chars at all).
if not re.search(LANGREGEX, txt):
return False
# Japanese only in the nameplate / control codes while the body is done.
return bool(re.search(LANGREGEX, _text_check_body(txt, entry.get("speaker_src", ""))))
def parseDocument(data, filename):
"""Translate every translatable leaf entry and return [data, tokens, error]."""
global PBAR
totalTokens = [0, 0]
entries = collectEntries(data)
is_names = data.get("kind") == "names"
def _translatable(e):
src = e.get("source")
if not (isinstance(src, str) and re.search(LANGREGEX, src)):
return False
# Already translated: look at ``text`` (source stays Japanese forever).
# Fresh extracts keep ``text == source``, so they still queue.
if IGNORETLTEXT and not _text_still_needs_translation(e):
return False
# names.json: only translate WolfDawn safe entries (per-name badge).
if is_names and not wolf_names.is_name_translatable(e):
return False
return True
# Only translate entries that still need work; names.json also requires a
# safe badge. Untouched leaves keep ``text == source`` so WolfDawn
# treats them as no-ops on inject.
translatable = [e for e in entries if _translatable(e)]
with tqdm(bar_format=BAR_FORMAT, position=POSITION, total=len(translatable), leave=LEAVE) as pbar:
pbar.desc = filename
PBAR = pbar
if translatable:
# Reshape first-line-speaker lines into the shared "[Speaker]: line"
# transport format. plans[i] carries what is needed to restore each
# entry after translation.
sources = []
plans = [] # (entry, prefix, has_speaker, is_firstline, code_map)
for entry in translatable:
src = entry["source"]
protected_src, code_map = wolf_codes.protect_wolf_codes(src)
is_firstline = entry.get("speaker_src", "") in wolf_speakers.FIRSTLINE_SRCS
split = wolf_speakers.split_source(
protected_src, entry.get("speaker_src", ""), SPEAKER_CONFIG
)
if split is not None:
prefix, speaker, body = split
sources.append(wolf_speakers.to_prefixed(speaker, body))
plans.append((entry, prefix, True, is_firstline, code_map))
else:
sources.append(protected_src)
plans.append((entry, "", False, is_firstline, code_map))
try:
response = translateAI(sources, [])
except Exception as e:
return [data, totalTokens, e]
translated, tokens = response[0], response[1]
totalTokens[0] += tokens[0]
totalTokens[1] += tokens[1]
# Write translations back. Skip estimate mode, and skip the batch
# collect pass (response is still the Japanese source list — wrapping
# it into ``text`` only reflows JP and is what poisoned translated/).
collecting = _batch_phase() == "collect"
if (
not ESTIMATE
and not collecting
and isinstance(translated, list)
and len(translated) == len(plans)
):
for (entry, prefix, has_speaker, is_firstline, code_map), text, src in zip(
plans, translated, sources
):
if not isinstance(text, str):
continue
# Model / collect echoed the sent payload — leave the entry alone.
if text == src:
continue
text = wolf_codes.restore_wolf_code_placeholders(text, code_map)
if has_speaker:
speaker_en, body_en = wolf_speakers.parse_prefixed(text)
if speaker_en is not None:
entry["text"] = wolf_speakers.restore_source(
prefix, speaker_en, body_en
)
else:
entry["text"] = prefix + text
else:
entry["text"] = text
wolf_codes.repair_entry(entry)
# Phase 0 feeds the glossary: harvest safe name values into vocab.txt
# (grouped by note) so the DB-text and dialogue phases keep item/skill/term
# names consistent. Runs even when every name was already translated (skip),
# so a resume still seeds vocab. Mirrors the RPGMaker DB-first strategy.
if is_names and not ESTIMATE:
_harvest_names_to_vocab(data)
# Foundation DB label fields (map names, titles) merge into vocab.txt without
# replacing entries already seeded from names.json.
elif data.get("kind") == "db" and not ESTIMATE:
_harvest_db_to_vocab(data)
return [data, totalTokens, None]
def _harvest_names_to_vocab(data):
"""Write short translated name labels into vocab.txt, grouped by note.
Profile blurbs and other content-shaped names are translated in names.json but
skipped here. Categories with no harvestable terms remove any stale section.
"""
try:
db_labels = wolf_names.derive_db_labels("files")
by_note: dict[str, list] = {}
touched_notes: set[str] = set()
for entry in data.get("names") or []:
if not wolf_names.is_name_translatable(entry):
continue
note = str(entry.get("note", ""))
touched_notes.add(note)
if not wolf_names.is_vocab_harvest_candidate(entry):
continue
src, dst = entry.get("source"), entry.get("text")
if not isinstance(src, str) or not isinstance(dst, str):
continue
by_note.setdefault(note, []).append((src, dst))
for note in touched_notes:
header = wolf_names.note_header(note, db_labels)
pairs = by_note.get(note, [])
if pairs:
wolf_vocab.update_vocab_section(header, pairs)
else:
wolf_vocab.remove_vocab_section(header)
except Exception:
traceback.print_exc()
def _harvest_db_to_vocab(data):
"""Merge short foundation DB labels into vocab.txt, grouped by sheet.
Uses merge mode so names.json-seeded entries for the same section stay put.
Descriptions and narrative dialogue are filtered out.
"""
try:
include_tiers, include_groups = wolf_db.load_db_filter_config()
by_sheet = wolf_db.collect_db_vocab_pairs(
data,
include_tiers=include_tiers,
include_groups=include_groups,
)
for type_name, pairs in by_sheet.items():
wolf_vocab.update_vocab_section(type_name, pairs, merge=True)
except Exception:
traceback.print_exc()
def getResultString(translatedData, translationTime, filename):
"""Format the per-file / total cost + status line for the console log."""
cost = calculateCost(translatedData[1][0], translatedData[1][1], MODEL)
totalTokenstring = (
Fore.YELLOW + "[Input: " + str(translatedData[1][0]) + "]"
"[Output: " + str(translatedData[1][1]) + "]"
"[Cost: ${:,.4f}".format(cost) + "]"
)
timeString = Fore.BLUE + "[" + str(round(translationTime, 1)) + "s]"
if translatedData[2] is None:
return filename + ": " + totalTokenstring + timeString + Fore.GREEN + " \u2713 " + Fore.RESET
try:
raise translatedData[2]
except Exception as e:
traceback.print_exc()
errorString = str(e) + Fore.RED
return (
filename + ": " + totalTokenstring + timeString
+ Fore.RED + " \u2717 " + errorString + Fore.RESET
)
def translateAI(text, history, history_ctx=None):
"""Thin wrapper around the shared translation entry point."""
global PBAR, MISMATCH, FILENAME
TRANSLATION_CONFIG.estimateMode = bool(ESTIMATE)
return sharedtranslateAI(
text=text,
history=history,
config=TRANSLATION_CONFIG,
filename=FILENAME,
pbar=PBAR,
lock=LOCK,
mismatchList=MISMATCH,
)